diff --git a/outputs/2025-11-24/12-15-43/.hydra/config.yaml b/outputs/2025-11-24/12-15-43/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..99a371108b39c87e5d97db54c3cdd2ab42787580
--- /dev/null
+++ b/outputs/2025-11-24/12-15-43/.hydra/config.yaml
@@ -0,0 +1,753 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest
+  experiment_name: sokoban-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: 100
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 2
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_frozen_lake_subagent_newenv_slippery_splitsingleturn
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-11-24/12-18-38/.hydra/overrides.yaml b/outputs/2025-11-24/12-18-38/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..df78aa5d804e1412811fd4f85e86c66d55b91756
--- /dev/null
+++ b/outputs/2025-11-24/12-18-38/.hydra/overrides.yaml
@@ -0,0 +1,7 @@
+- system.CUDA_VISIBLE_DEVICES=2
+- trainer.experiment_name=sokoban-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-24/14-30-33/.hydra/config.yaml b/outputs/2025-11-24/14-30-33/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3c3e0cef2141ba1392e3e5ff610bfe9130d2fd0b
--- /dev/null
+++ b/outputs/2025-11-24/14-30-33/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: 100
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-11-24/17-05-35/.hydra/hydra.yaml b/outputs/2025-11-24/17-05-35/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8bfbd6af886fefe11bdf8f28d831a0fa81e5f59f
--- /dev/null
+++ b/outputs/2025-11-24/17-05-35/.hydra/hydra.yaml
@@ -0,0 +1,176 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=2
+    - trainer.experiment_name=sokoban-ppo
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,system.CUDA_VISIBLE_DEVICES=2,trainer.experiment_name=sokoban-ppo
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-11-24/17-05-35
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-11-27/22-21-59/.hydra/overrides.yaml b/outputs/2025-11-27/22-21-59/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2c489d70e1de695a91635f9027b5be0a9e7e962d
--- /dev/null
+++ b/outputs/2025-11-27/22-21-59/.hydra/overrides.yaml
@@ -0,0 +1,7 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/01-07-11/.hydra/overrides.yaml b/outputs/2025-11-28/01-07-11/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..05bdf3e48ccdd844c0d515bd01f49544bc1aaef5
--- /dev/null
+++ b/outputs/2025-11-28/01-07-11/.hydra/overrides.yaml
@@ -0,0 +1,7 @@
+- system.CUDA_VISIBLE_DEVICES=2
+- trainer.experiment_name=frozen_lake-ppo-multitask-slippery
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/01-30-10/.hydra/overrides.yaml b/outputs/2025-11-28/01-30-10/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..82267f100340837444266dff04347c7342497f3c
--- /dev/null
+++ b/outputs/2025-11-28/01-30-10/.hydra/overrides.yaml
@@ -0,0 +1,8 @@
+- system.CUDA_VISIBLE_DEVICES=3
+- trainer.experiment_name=bandit-ppo-multitask
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
+ilter_ratio=1
diff --git a/outputs/2025-11-28/01-30-10/train.log b/outputs/2025-11-28/01-30-10/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-11-28/02-19-11/.hydra/overrides.yaml b/outputs/2025-11-28/02-19-11/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..05bdf3e48ccdd844c0d515bd01f49544bc1aaef5
--- /dev/null
+++ b/outputs/2025-11-28/02-19-11/.hydra/overrides.yaml
@@ -0,0 +1,7 @@
+- system.CUDA_VISIBLE_DEVICES=2
+- trainer.experiment_name=frozen_lake-ppo-multitask-slippery
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/10-45-12/.hydra/hydra.yaml b/outputs/2025-11-28/10-45-12/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..08ae37d107e47d2d06a9e69e9280caaa02bd21d1
--- /dev/null
+++ b/outputs/2025-11-28/10-45-12/.hydra/hydra.yaml
@@ -0,0 +1,176 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=3
+    - trainer.experiment_name=bandit-ppo-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,system.CUDA_VISIBLE_DEVICES=3,trainer.experiment_name=bandit-ppo-frommlp
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-11-28/10-45-12
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-11-28/10-45-12/train.log b/outputs/2025-11-28/10-45-12/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-11-28/11-13-34/.hydra/overrides.yaml b/outputs/2025-11-28/11-13-34/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7b01644feb05a86becb7797f5b3d601ffba97175
--- /dev/null
+++ b/outputs/2025-11-28/11-13-34/.hydra/overrides.yaml
@@ -0,0 +1,7 @@
+- system.CUDA_VISIBLE_DEVICES=3
+- trainer.experiment_name=bandit-ppo-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/14-19-49/.hydra/config.yaml b/outputs/2025-11-28/14-19-49/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ceb23dd4821239f06d7d92393a587f84cbb1f8e1
--- /dev/null
+++ b/outputs/2025-11-28/14-19-49/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: 100
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-11-28/14-19-49/.hydra/overrides.yaml b/outputs/2025-11-28/14-19-49/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2c489d70e1de695a91635f9027b5be0a9e7e962d
--- /dev/null
+++ b/outputs/2025-11-28/14-19-49/.hydra/overrides.yaml
@@ -0,0 +1,7 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/14-24-14/.hydra/config.yaml b/outputs/2025-11-28/14-24-14/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e6ac7c1a47b92700525396f0067cc5c57c773563
--- /dev/null
+++ b/outputs/2025-11-28/14-24-14/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 100
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0,1
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-11-28/14-24-14/.hydra/overrides.yaml b/outputs/2025-11-28/14-24-14/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6ce7e01b3646aca0e308bde77bc3e99244ec6cae
--- /dev/null
+++ b/outputs/2025-11-28/14-24-14/.hydra/overrides.yaml
@@ -0,0 +1,8 @@
+- system.CUDA_VISIBLE_DEVICES='0,1'
+- trainer.n_gpus_per_node=2
+- trainer.experiment_name=bandit-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/14-24-14/train.log b/outputs/2025-11-28/14-24-14/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-11-28/14-26-21/.hydra/overrides.yaml b/outputs/2025-11-28/14-26-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6ce7e01b3646aca0e308bde77bc3e99244ec6cae
--- /dev/null
+++ b/outputs/2025-11-28/14-26-21/.hydra/overrides.yaml
@@ -0,0 +1,8 @@
+- system.CUDA_VISIBLE_DEVICES='0,1'
+- trainer.n_gpus_per_node=2
+- trainer.experiment_name=bandit-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/16-14-49/.hydra/config.yaml b/outputs/2025-11-28/16-14-49/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..772144fe1297a77a2a3f0cdb235672136fd51697
--- /dev/null
+++ b/outputs/2025-11-28/16-14-49/.hydra/config.yaml
@@ -0,0 +1,753 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sokoban-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 100
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-11-28/16-14-49/.hydra/hydra.yaml b/outputs/2025-11-28/16-14-49/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ce8146317d9f7c53c18cb368b366c8b5e86f4dc9
--- /dev/null
+++ b/outputs/2025-11-28/16-14-49/.hydra/hydra.yaml
@@ -0,0 +1,177 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2,3'
+    - trainer.n_gpus_per_node=2
+    - trainer.experiment_name=sokoban-ppo
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,system.CUDA_VISIBLE_DEVICES='2,3',trainer.experiment_name=sokoban-ppo,trainer.n_gpus_per_node=2
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-11-28/16-14-49
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-11-28/16-14-50/.hydra/config.yaml b/outputs/2025-11-28/16-14-50/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8df154ece496c6acbfbadca9bcef40f9fbedb728
--- /dev/null
+++ b/outputs/2025-11-28/16-14-50/.hydra/config.yaml
@@ -0,0 +1,753 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: frozen_lake-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 100
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 4,5
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-11-28/16-14-50/.hydra/hydra.yaml b/outputs/2025-11-28/16-14-50/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e75bf476794b8490836844e3b7d80347f4abd1c3
--- /dev/null
+++ b/outputs/2025-11-28/16-14-50/.hydra/hydra.yaml
@@ -0,0 +1,177 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5'
+    - trainer.n_gpus_per_node=2
+    - trainer.experiment_name=frozen_lake-ppo
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,system.CUDA_VISIBLE_DEVICES='4,5',trainer.experiment_name=frozen_lake-ppo,trainer.n_gpus_per_node=2
+    id: ???
+    num: ???
+    config_name: _3_frozen_lake
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-11-28/16-14-50
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-11-28/16-14-50/.hydra/overrides.yaml b/outputs/2025-11-28/16-14-50/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..33da5949eb50839c857562b221304607233f4f27
--- /dev/null
+++ b/outputs/2025-11-28/16-14-50/.hydra/overrides.yaml
@@ -0,0 +1,8 @@
+- system.CUDA_VISIBLE_DEVICES='4,5'
+- trainer.n_gpus_per_node=2
+- trainer.experiment_name=frozen_lake-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/16-20-19/.hydra/config.yaml b/outputs/2025-11-28/16-20-19/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cc41213dff6405e4613e152b29e358d97cb30786
--- /dev/null
+++ b/outputs/2025-11-28/16-20-19/.hydra/config.yaml
@@ -0,0 +1,754 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sokoban-ppo-box1
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: 100
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+ kwargs: null
diff --git a/outputs/2025-11-28/16-20-19/.hydra/overrides.yaml b/outputs/2025-11-28/16-20-19/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2a80bbb860b3cbf76a04725d403f6dab3829d60c
--- /dev/null
+++ b/outputs/2025-11-28/16-20-19/.hydra/overrides.yaml
@@ -0,0 +1,8 @@
+- system.CUDA_VISIBLE_DEVICES='4,5'
+- trainer.n_gpus_per_node=2
+- trainer.experiment_name=frozen_lake-ppo-slippery
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-11-28/16-20-19/train.log b/outputs/2025-11-28/16-20-19/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-11-28/16-58-54/.hydra/hydra.yaml b/outputs/2025-11-28/16-58-54/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..056e53281b4c893298b03191c80f85e1d329deed
--- /dev/null
+++ b/outputs/2025-11-28/16-58-54/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2,3'
+    - trainer.n_gpus_per_node=2
+    - trainer.experiment_name=sokoban-ppo-box1
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,system.CUDA_VISIBLE_DEVICES='2,3',trainer.experiment_name=sokoban-ppo-box1,trainer.n_gpus_per_node=2
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-11-28/16-58-54
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
+lt
+  verbose: false
diff --git a/outputs/2025-11-28/16-58-54/.hydra/overrides.yaml b/outputs/2025-11-28/16-58-54/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c3e592b9a949e70b41e7edc76838033061b293a7
--- /dev/null
+++ b/outputs/2025-11-28/16-58-54/.hydra/overrides.yaml
@@ -0,0 +1,9 @@
+- system.CUDA_VISIBLE_DEVICES='2,3'
+- trainer.n_gpus_per_node=2
+- trainer.experiment_name=sokoban-ppo-box1
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
+ratio=1
diff --git a/outputs/2025-12-02/00-06-34/.hydra/config.yaml b/outputs/2025-12-02/00-06-34/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..928bc77fcc84835b2fc44d26bb96952ffbd9be9a
--- /dev/null
+++ b/outputs/2025-12-02/00-06-34/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: bandit-ppo-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_bandit
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/00-06-34/.hydra/overrides.yaml b/outputs/2025-12-02/00-06-34/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..65005e83bf34670c1a4fc564fc354da7c29daf82
--- /dev/null
+++ b/outputs/2025-12-02/00-06-34/.hydra/overrides.yaml
@@ -0,0 +1,3 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.n_gpus_per_node=1
+- trainer.experiment_name=bandit-ppo-frommlp
diff --git a/outputs/2025-12-02/00-50-08/.hydra/config.yaml b/outputs/2025-12-02/00-50-08/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..928bc77fcc84835b2fc44d26bb96952ffbd9be9a
--- /dev/null
+++ b/outputs/2025-12-02/00-50-08/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: bandit-ppo-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_bandit
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/00-50-08/.hydra/hydra.yaml b/outputs/2025-12-02/00-50-08/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d0f786a44559e3336637ff4f28ad5f0196bb8434
--- /dev/null
+++ b/outputs/2025-12-02/00-50-08/.hydra/hydra.yaml
@@ -0,0 +1,172 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.n_gpus_per_node=1
+    - trainer.experiment_name=bandit-ppo-frommlp
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=bandit-ppo-frommlp,trainer.n_gpus_per_node=1
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/00-50-08
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/00-50-08/train.log b/outputs/2025-12-02/00-50-08/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/09-21-16/.hydra/config.yaml b/outputs/2025-12-02/09-21-16/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dc4011c3aa5ba2099bcdbe3bb487868aaf0bf90e
--- /dev/null
+++ b/outputs/2025-12-02/09-21-16/.hydra/config.yaml
@@ -0,0 +1,753 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-ppo-box2
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/09-21-16/.hydra/hydra.yaml b/outputs/2025-12-02/09-21-16/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0ca2a5bcb65790b87d72baada0716ee5b164495b
--- /dev/null
+++ b/outputs/2025-12-02/09-21-16/.hydra/hydra.yaml
@@ -0,0 +1,178 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2,3'
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+    - trainer.experiment_name=sokoban-ppo-box2
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask,system.CUDA_VISIBLE_DEVICES='2,3',trainer.experiment_name=sokoban-ppo-box2,trainer.n_gpus_per_node=2
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/09-21-16
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/09-21-16/.hydra/overrides.yaml b/outputs/2025-12-02/09-21-16/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..87b79fdae0a221bc40e33b28fc6860c20bfa4065
--- /dev/null
+++ b/outputs/2025-12-02/09-21-16/.hydra/overrides.yaml
@@ -0,0 +1,9 @@
+- system.CUDA_VISIBLE_DEVICES='2,3'
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+- trainer.experiment_name=sokoban-ppo-box2
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-02/09-21-16/train.log b/outputs/2025-12-02/09-21-16/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/09-26-07/.hydra/config.yaml b/outputs/2025-12-02/09-26-07/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fcb6955b1a6ad74e3849470754739391fd5f1137
--- /dev/null
+++ b/outputs/2025-12-02/09-26-07/.hydra/config.yaml
@@ -0,0 +1,754 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-ppo-box2
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+   kwargs: null
diff --git a/outputs/2025-12-02/09-26-07/.hydra/hydra.yaml b/outputs/2025-12-02/09-26-07/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fde9da3e214764d5457ecd279e7283643016defd
--- /dev/null
+++ b/outputs/2025-12-02/09-26-07/.hydra/hydra.yaml
@@ -0,0 +1,178 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2,3'
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+    - trainer.experiment_name=sokoban-ppo-box2
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask,system.CUDA_VISIBLE_DEVICES='2,3',trainer.experiment_name=sokoban-ppo-box2,trainer.n_gpus_per_node=2
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/09-26-07
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/09-26-07/.hydra/overrides.yaml b/outputs/2025-12-02/09-26-07/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..87b79fdae0a221bc40e33b28fc6860c20bfa4065
--- /dev/null
+++ b/outputs/2025-12-02/09-26-07/.hydra/overrides.yaml
@@ -0,0 +1,9 @@
+- system.CUDA_VISIBLE_DEVICES='2,3'
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+- trainer.experiment_name=sokoban-ppo-box2
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-02/09-26-07/train.log b/outputs/2025-12-02/09-26-07/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-20-22/.hydra/config.yaml b/outputs/2025-12-02/10-20-22/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..075235a278c79883510842fe708f6d6e160716ad
--- /dev/null
+++ b/outputs/2025-12-02/10-20-22/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0,1
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/10-20-22/.hydra/hydra.yaml b/outputs/2025-12-02/10-20-22/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0e702caf1467f5facadf4bc5c4475f8dd3c9f1c0
--- /dev/null
+++ b/outputs/2025-12-02/10-20-22/.hydra/hydra.yaml
@@ -0,0 +1,178 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1'
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+    - trainer.experiment_name=bandit-ppo
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask,system.CUDA_VISIBLE_DEVICES='0,1',trainer.experiment_name=bandit-ppo,trainer.n_gpus_per_node=2
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-20-22
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-20-22/.hydra/overrides.yaml b/outputs/2025-12-02/10-20-22/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cd636a171016689ea31749641d807fbf3eba3525
--- /dev/null
+++ b/outputs/2025-12-02/10-20-22/.hydra/overrides.yaml
@@ -0,0 +1,9 @@
+- system.CUDA_VISIBLE_DEVICES='0,1'
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+- trainer.experiment_name=bandit-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-02/10-28-44/.hydra/config.yaml b/outputs/2025-12-02/10-28-44/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f63b9dce0be052ddaeba2bfae4c620775ae33d43
--- /dev/null
+++ b/outputs/2025-12-02/10-28-44/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: null
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/10-28-44/.hydra/hydra.yaml b/outputs/2025-12-02/10-28-44/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..08e54b229cf479d0d57bcbc826a86a0903a2cf12
--- /dev/null
+++ b/outputs/2025-12-02/10-28-44/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-28-44
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-28-44/.hydra/overrides.yaml b/outputs/2025-12-02/10-28-44/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/10-28-44/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/10-28-44/train.log b/outputs/2025-12-02/10-28-44/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-29-10/.hydra/config.yaml b/outputs/2025-12-02/10-29-10/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..138ae611a858662404c1430bf6beaaef3f7cf5f2
--- /dev/null
+++ b/outputs/2025-12-02/10-29-10/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/10-29-10/.hydra/hydra.yaml b/outputs/2025-12-02/10-29-10/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bd4b0512ff38145de6f4fa286777adbb42277112
--- /dev/null
+++ b/outputs/2025-12-02/10-29-10/.hydra/hydra.yaml
@@ -0,0 +1,172 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+  job:
+    name: train
+    chdir: null
+    override_dirname: model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask,system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-29-10
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-29-10/train.log b/outputs/2025-12-02/10-29-10/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-30-01/.hydra/overrides.yaml b/outputs/2025-12-02/10-30-01/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..70a5731b7b3e9a0ba410c07cc2e16ba08e560817
--- /dev/null
+++ b/outputs/2025-12-02/10-30-01/.hydra/overrides.yaml
@@ -0,0 +1,3 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
diff --git a/outputs/2025-12-02/10-30-01/train.log b/outputs/2025-12-02/10-30-01/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-31-16/.hydra/config.yaml b/outputs/2025-12-02/10-31-16/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/10-31-16/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/10-31-16/.hydra/hydra.yaml b/outputs/2025-12-02/10-31-16/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9fcd1399be05dc52382e2d5427ef624f4de9018f
--- /dev/null
+++ b/outputs/2025-12-02/10-31-16/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-31-16
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-31-16/.hydra/overrides.yaml b/outputs/2025-12-02/10-31-16/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/10-31-16/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/10-31-16/train.log b/outputs/2025-12-02/10-31-16/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-36-25/.hydra/config.yaml b/outputs/2025-12-02/10-36-25/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/10-36-25/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/10-36-25/.hydra/hydra.yaml b/outputs/2025-12-02/10-36-25/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6e3c1259c1fe4252dc905b29913cd30b954b678f
--- /dev/null
+++ b/outputs/2025-12-02/10-36-25/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-36-25
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-36-25/.hydra/overrides.yaml b/outputs/2025-12-02/10-36-25/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/10-36-25/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/10-36-25/train.log b/outputs/2025-12-02/10-36-25/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-38-36/.hydra/hydra.yaml b/outputs/2025-12-02/10-38-36/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4ec7d7642552d5faf9edad1600360707cfbb7bd2
--- /dev/null
+++ b/outputs/2025-12-02/10-38-36/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1'
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+    - trainer.experiment_name=bandit-ppo
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask,system.CUDA_VISIBLE_DEVICES='0,1',trainer.experiment_name=bandit-ppo,trainer.n_gpus_per_node=2
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-38-36
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
+dra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-38-36/.hydra/overrides.yaml b/outputs/2025-12-02/10-38-36/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1d8738662e0865c75cef7d6e977e20bd81f45ea7
--- /dev/null
+++ b/outputs/2025-12-02/10-38-36/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1'
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+- trainer.experiment_name=bandit-ppo
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
+_filter_ratio=1
diff --git a/outputs/2025-12-02/10-38-36/train.log b/outputs/2025-12-02/10-38-36/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-42-50/.hydra/config.yaml b/outputs/2025-12-02/10-42-50/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/10-42-50/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/10-42-50/.hydra/hydra.yaml b/outputs/2025-12-02/10-42-50/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..83ce74811d0447d69aee1f66d8a3c80cef926b5f
--- /dev/null
+++ b/outputs/2025-12-02/10-42-50/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-42-50
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-42-50/train.log b/outputs/2025-12-02/10-42-50/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-44-30/.hydra/hydra.yaml b/outputs/2025-12-02/10-44-30/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c82ae44bf3cadb245e200938af455b3fdb8efa8e
--- /dev/null
+++ b/outputs/2025-12-02/10-44-30/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-44-30
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-44-30/.hydra/overrides.yaml b/outputs/2025-12-02/10-44-30/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/10-44-30/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/10-44-30/train.log b/outputs/2025-12-02/10-44-30/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-46-43/.hydra/hydra.yaml b/outputs/2025-12-02/10-46-43/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9bf690ca30fe892a33db11066b4eb665fb9d0ced
--- /dev/null
+++ b/outputs/2025-12-02/10-46-43/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-46-43
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-46-43/.hydra/overrides.yaml b/outputs/2025-12-02/10-46-43/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/10-46-43/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/10-46-43/train.log b/outputs/2025-12-02/10-46-43/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-49-53/.hydra/hydra.yaml b/outputs/2025-12-02/10-49-53/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..755d409d68ee3cc59e3075cd3b466b361aa7b156
--- /dev/null
+++ b/outputs/2025-12-02/10-49-53/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/10-49-53
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/10-49-53/.hydra/overrides.yaml b/outputs/2025-12-02/10-49-53/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/10-49-53/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/10-49-53/train.log b/outputs/2025-12-02/10-49-53/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/10-53-37/.hydra/config.yaml b/outputs/2025-12-02/10-53-37/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/10-53-37/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/10-53-37/.hydra/overrides.yaml b/outputs/2025-12-02/10-53-37/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/10-53-37/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/11-13-16/.hydra/config.yaml b/outputs/2025-12-02/11-13-16/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/11-13-16/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/11-13-16/.hydra/hydra.yaml b/outputs/2025-12-02/11-13-16/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..540297dbb68ffb12231305ca41dc05f6dbbacc67
--- /dev/null
+++ b/outputs/2025-12-02/11-13-16/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/11-13-16
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/11-14-37/.hydra/config.yaml b/outputs/2025-12-02/11-14-37/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/11-14-37/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/11-14-37/.hydra/overrides.yaml b/outputs/2025-12-02/11-14-37/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/11-14-37/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/11-14-37/train.log b/outputs/2025-12-02/11-14-37/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/11-17-41/.hydra/config.yaml b/outputs/2025-12-02/11-17-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/11-17-41/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/11-17-41/train.log b/outputs/2025-12-02/11-17-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/11-20-28/.hydra/config.yaml b/outputs/2025-12-02/11-20-28/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/11-20-28/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/11-20-28/.hydra/hydra.yaml b/outputs/2025-12-02/11-20-28/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3bcbd5c25a189f15ccb9672cef87cdc9c861fce9
--- /dev/null
+++ b/outputs/2025-12-02/11-20-28/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/11-20-28
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/11-20-28/train.log b/outputs/2025-12-02/11-20-28/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/11-22-01/.hydra/config.yaml b/outputs/2025-12-02/11-22-01/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48840bf5b7fcf91b28480ab27bfa907f8ef39cc0
--- /dev/null
+++ b/outputs/2025-12-02/11-22-01/.hydra/config.yaml
@@ -0,0 +1,755 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-ppo
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-02/11-22-01/.hydra/hydra.yaml b/outputs/2025-12-02/11-22-01/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..32a821f6acaa81d9531f8f0add0ee45114e19f26
--- /dev/null
+++ b/outputs/2025-12-02/11-22-01/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/11-22-01
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/11-22-01/.hydra/overrides.yaml b/outputs/2025-12-02/11-22-01/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/11-22-01/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/11-43-57/.hydra/hydra.yaml b/outputs/2025-12-02/11-43-57/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5dc912988d1499371819e167fb015946864e0d7b
--- /dev/null
+++ b/outputs/2025-12-02/11-43-57/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES=0
+    - trainer.experiment_name=bandit-ppo
+  job:
+    name: train
+    chdir: null
+    override_dirname: system.CUDA_VISIBLE_DEVICES=0,trainer.experiment_name=bandit-ppo
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-02/11-43-57
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-02/11-43-57/train.log b/outputs/2025-12-02/11-43-57/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/11-45-48/.hydra/overrides.yaml b/outputs/2025-12-02/11-45-48/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..39bb5871ab345f104c0f9af211f19552dc5ad0cf
--- /dev/null
+++ b/outputs/2025-12-02/11-45-48/.hydra/overrides.yaml
@@ -0,0 +1,2 @@
+- system.CUDA_VISIBLE_DEVICES=0
+- trainer.experiment_name=bandit-ppo
diff --git a/outputs/2025-12-02/11-45-48/train.log b/outputs/2025-12-02/11-45-48/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/11-50-18/train.log b/outputs/2025-12-02/11-50-18/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/13-28-10/train.log b/outputs/2025-12-02/13-28-10/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/15-07-06/train.log b/outputs/2025-12-02/15-07-06/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/17-11-10/train.log b/outputs/2025-12-02/17-11-10/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-02/19-45-35/train.log b/outputs/2025-12-02/19-45-35/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-03/15-40-21/.hydra/config.yaml b/outputs/2025-12-03/15-40-21/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d3422e5f6eb7aa3344ce8556f09c55a9b2d33146
--- /dev/null
+++ b/outputs/2025-12-03/15-40-21/.hydra/config.yaml
@@ -0,0 +1,753 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: frozen_lake-ppo-slippery-multitask
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_multitask
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-05/11-00-42/.hydra/config.yaml b/outputs/2025-12-05/11-00-42/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4c32748f43afc4af9fcd6bcfb27e7849694ec600
--- /dev/null
+++ b/outputs/2025-12-05/11-00-42/.hydra/config.yaml
@@ -0,0 +1,780 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: 2048
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  Game2048:
+    env_type: game_2048
+    max_actions_per_traj: 5
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  Blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Game2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - Game2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-05/11-00-42/.hydra/hydra.yaml b/outputs/2025-12-05/11-00-42/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..09b3fac006ad6cfb6a56b3f4310c785dbddc1dad
--- /dev/null
+++ b/outputs/2025-12-05/11-00-42/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=2048
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=2048,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-05/11-00-42
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-06/20-15-18/train.log b/outputs/2025-12-06/20-15-18/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-06/23-42-39/.hydra/config.yaml b/outputs/2025-12-06/23-42-39/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e1211815c328d96fb4f7ea97ca496c694a81087e
--- /dev/null
+++ b/outputs/2025-12-06/23-42-39/.hydra/config.yaml
@@ -0,0 +1,780 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: blackjack-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_blackjack
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - blackjack
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - blackjack
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-06/23-45-22/.hydra/hydra.yaml b/outputs/2025-12-06/23-45-22/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b4121f7a0cc9f4632f0e0dc657fcf6202d18011c
--- /dev/null
+++ b/outputs/2025-12-06/23-45-22/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_blackjack
+    - trainer.experiment_name=blackjack-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_blackjack,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=blackjack-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _9_blackjack
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-06/23-45-22
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
+false
diff --git a/outputs/2025-12-07/05-45-49/.hydra/overrides.yaml b/outputs/2025-12-07/05-45-49/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..257c3f117793a9d242bbb09edb4cfd690bf82794
--- /dev/null
+++ b/outputs/2025-12-07/05-45-49/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_blackjack
+- trainer.experiment_name=blackjack-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/11-33-31/.hydra/hydra.yaml b/outputs/2025-12-07/11-33-31/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d6a64c46fd6ec4a66c2b63d7616d3574b0dec9fc
--- /dev/null
+++ b/outputs/2025-12-07/11-33-31/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/11-33-31
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/11-46-18/.hydra/config.yaml b/outputs/2025-12-07/11-46-18/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2eacee8b1cc82286bd497841a034ef34f95dfc55
--- /dev/null
+++ b/outputs/2025-12-07/11-46-18/.hydra/config.yaml
@@ -0,0 +1,799 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore each face to a single color.
+
+      Available actions are face rotations: U, U'', D, D'', L, L'', R, R'', F, F'',
+      B, B''.
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U || R || F''</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 5
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/11-46-18/.hydra/hydra.yaml b/outputs/2025-12-07/11-46-18/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d0175417abf5c0b86d0f99ca101c31317f587c40
--- /dev/null
+++ b/outputs/2025-12-07/11-46-18/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/11-46-18
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/11-46-18/.hydra/overrides.yaml b/outputs/2025-12-07/11-46-18/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..074482467edc64a15411978fd0a1b97591a95613
--- /dev/null
+++ b/outputs/2025-12-07/11-46-18/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/11-50-15/.hydra/config.yaml b/outputs/2025-12-07/11-50-15/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2eacee8b1cc82286bd497841a034ef34f95dfc55
--- /dev/null
+++ b/outputs/2025-12-07/11-50-15/.hydra/config.yaml
@@ -0,0 +1,799 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore each face to a single color.
+
+      Available actions are face rotations: U, U'', D, D'', L, L'', R, R'', F, F'',
+      B, B''.
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U || R || F''</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 5
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/11-50-15/.hydra/hydra.yaml b/outputs/2025-12-07/11-50-15/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d812e5a14fdc3daf70c028ea08ac0c06c634d116
--- /dev/null
+++ b/outputs/2025-12-07/11-50-15/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/11-50-15
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/12-09-21/.hydra/config.yaml b/outputs/2025-12-07/12-09-21/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bebd35c1ee6b8cc0d569bba0bb8410d76cb13d41
--- /dev/null
+++ b/outputs/2025-12-07/12-09-21/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 5
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/12-09-21/.hydra/hydra.yaml b/outputs/2025-12-07/12-09-21/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6b16d2e7674eb953a3c78bd39c54822c2d71b979
--- /dev/null
+++ b/outputs/2025-12-07/12-09-21/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/12-09-21
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/12-09-21/.hydra/overrides.yaml b/outputs/2025-12-07/12-09-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..074482467edc64a15411978fd0a1b97591a95613
--- /dev/null
+++ b/outputs/2025-12-07/12-09-21/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/12-09-21/train.log b/outputs/2025-12-07/12-09-21/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-07/12-15-28/.hydra/config.yaml b/outputs/2025-12-07/12-15-28/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bebd35c1ee6b8cc0d569bba0bb8410d76cb13d41
--- /dev/null
+++ b/outputs/2025-12-07/12-15-28/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 5
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/12-15-28/.hydra/hydra.yaml b/outputs/2025-12-07/12-15-28/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8ecbbb3acc902a49ad9dc32694998a7f5a9253bb
--- /dev/null
+++ b/outputs/2025-12-07/12-15-28/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/12-15-28
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/12-15-28/.hydra/overrides.yaml b/outputs/2025-12-07/12-15-28/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..074482467edc64a15411978fd0a1b97591a95613
--- /dev/null
+++ b/outputs/2025-12-07/12-15-28/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/12-17-44/.hydra/config.yaml b/outputs/2025-12-07/12-17-44/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bebd35c1ee6b8cc0d569bba0bb8410d76cb13d41
--- /dev/null
+++ b/outputs/2025-12-07/12-17-44/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 5
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/12-17-44/.hydra/hydra.yaml b/outputs/2025-12-07/12-17-44/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4a5e86c995a8ff22a3647ae31d31fde3e3a7b07d
--- /dev/null
+++ b/outputs/2025-12-07/12-17-44/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/12-17-44
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/12-17-44/.hydra/overrides.yaml b/outputs/2025-12-07/12-17-44/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..074482467edc64a15411978fd0a1b97591a95613
--- /dev/null
+++ b/outputs/2025-12-07/12-17-44/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/13-32-20/.hydra/config.yaml b/outputs/2025-12-07/13-32-20/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bebd35c1ee6b8cc0d569bba0bb8410d76cb13d41
--- /dev/null
+++ b/outputs/2025-12-07/13-32-20/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 5
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/13-32-20/.hydra/hydra.yaml b/outputs/2025-12-07/13-32-20/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9bfa3abd7ba5165eb57e07da1f6bbebc959be162
--- /dev/null
+++ b/outputs/2025-12-07/13-32-20/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/13-32-20
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/13-32-20/.hydra/overrides.yaml b/outputs/2025-12-07/13-32-20/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..074482467edc64a15411978fd0a1b97591a95613
--- /dev/null
+++ b/outputs/2025-12-07/13-32-20/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/13-32-20/train.log b/outputs/2025-12-07/13-32-20/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-07/15-19-07/.hydra/config.yaml b/outputs/2025-12-07/15-19-07/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8ed44712315daeaba74ec28e53be85c8d4327ffb
--- /dev/null
+++ b/outputs/2025-12-07/15-19-07/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 5
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/15-19-07/.hydra/hydra.yaml b/outputs/2025-12-07/15-19-07/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..be8158ee7e175786f5f2bff906129d743e366f92
--- /dev/null
+++ b/outputs/2025-12-07/15-19-07/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube
+    - trainer.experiment_name=rubikscube-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/15-19-07
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/15-19-07/.hydra/overrides.yaml b/outputs/2025-12-07/15-19-07/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a80ed67f273f2c0116f76aed0cf264b6db05b525
--- /dev/null
+++ b/outputs/2025-12-07/15-19-07/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube
+- trainer.experiment_name=rubikscube-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/15-55-18/.hydra/hydra.yaml b/outputs/2025-12-07/15-55-18/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b4ab640c21dade6b425cdf18c1e65d37b811b0fa
--- /dev/null
+++ b/outputs/2025-12-07/15-55-18/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube
+    - trainer.experiment_name=rubikscube-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/15-55-18
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/15-55-18/.hydra/overrides.yaml b/outputs/2025-12-07/15-55-18/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a80ed67f273f2c0116f76aed0cf264b6db05b525
--- /dev/null
+++ b/outputs/2025-12-07/15-55-18/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube
+- trainer.experiment_name=rubikscube-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/16-13-03/.hydra/config.yaml b/outputs/2025-12-07/16-13-03/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..05e2e60a38740aabee22576a421d4d16f585daa9
--- /dev/null
+++ b/outputs/2025-12-07/16-13-03/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/16-13-03/.hydra/hydra.yaml b/outputs/2025-12-07/16-13-03/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..17b5a387ea7e49f5b50c371e1f4d41345a43de3e
--- /dev/null
+++ b/outputs/2025-12-07/16-13-03/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/16-13-03
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/16-13-03/.hydra/overrides.yaml b/outputs/2025-12-07/16-13-03/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..074482467edc64a15411978fd0a1b97591a95613
--- /dev/null
+++ b/outputs/2025-12-07/16-13-03/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/16-13-03/train.log b/outputs/2025-12-07/16-13-03/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-07/18-02-27/.hydra/config.yaml b/outputs/2025-12-07/18-02-27/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..148e6229700878a07e0cd7025cb17326cda82a3c
--- /dev/null
+++ b/outputs/2025-12-07/18-02-27/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/18-02-27/.hydra/hydra.yaml b/outputs/2025-12-07/18-02-27/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d6c018250240acf588bd7efbbc8211937ee9b032
--- /dev/null
+++ b/outputs/2025-12-07/18-02-27/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/18-02-27
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/18-02-27/.hydra/overrides.yaml b/outputs/2025-12-07/18-02-27/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..074482467edc64a15411978fd0a1b97591a95613
--- /dev/null
+++ b/outputs/2025-12-07/18-02-27/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-07/18-02-27/train.log b/outputs/2025-12-07/18-02-27/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-07/21-59-48/.hydra/config.yaml b/outputs/2025-12-07/21-59-48/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..44a3a809f2350d3ac41ab3f4fceb92c620f7f214
--- /dev/null
+++ b/outputs/2025-12-07/21-59-48/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-1
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-07/21-59-48/.hydra/hydra.yaml b/outputs/2025-12-07/21-59-48/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..478d0418c8932818d1d099729c2f1354b5333ee7
--- /dev/null
+++ b/outputs/2025-12-07/21-59-48/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=rubikscube-1
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-1,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-07/21-59-48
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-07/21-59-48/.hydra/overrides.yaml b/outputs/2025-12-07/21-59-48/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1d23e7cc19660126391ad1564f6d9f1601ad2b47
--- /dev/null
+++ b/outputs/2025-12-07/21-59-48/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=rubikscube-1
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/00-34-18/.hydra/config.yaml b/outputs/2025-12-08/00-34-18/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0b33c803d63e1e57502f03c728aa2dd540e0d003
--- /dev/null
+++ b/outputs/2025-12-08/00-34-18/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-1
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/00-34-18/.hydra/hydra.yaml b/outputs/2025-12-08/00-34-18/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e196e093b1be1c3bd111b53fc8a2c42eb7b83097
--- /dev/null
+++ b/outputs/2025-12-08/00-34-18/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=rubikscube-1
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube-1,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/00-34-18
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/00-34-18/.hydra/overrides.yaml b/outputs/2025-12-08/00-34-18/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3702609e0119fab881c46370c3e0f88d336a4f0f
--- /dev/null
+++ b/outputs/2025-12-08/00-34-18/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=rubikscube-1
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/00-34-18/train.log b/outputs/2025-12-08/00-34-18/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/04-39-51/.hydra/config.yaml b/outputs/2025-12-08/04-39-51/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0b33c803d63e1e57502f03c728aa2dd540e0d003
--- /dev/null
+++ b/outputs/2025-12-08/04-39-51/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-1
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/04-39-51/.hydra/hydra.yaml b/outputs/2025-12-08/04-39-51/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..83da7111188f1dcde69913f0625c39c1aa8f9e64
--- /dev/null
+++ b/outputs/2025-12-08/04-39-51/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=rubikscube-1
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube-1,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/04-39-51
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/04-39-51/.hydra/overrides.yaml b/outputs/2025-12-08/04-39-51/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3702609e0119fab881c46370c3e0f88d336a4f0f
--- /dev/null
+++ b/outputs/2025-12-08/04-39-51/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=rubikscube-1
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/09-09-15/.hydra/config.yaml b/outputs/2025-12-08/09-09-15/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0c8a8e62c9de66cff22bffc7a98d2d45f3d948ea
--- /dev/null
+++ b/outputs/2025-12-08/09-09-15/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/09-09-15/.hydra/hydra.yaml b/outputs/2025-12-08/09-09-15/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..03f8a18e2007f6d66fc06f86a6047dd7970a8f03
--- /dev/null
+++ b/outputs/2025-12-08/09-09-15/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube-1
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube-1,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/09-09-15
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/09-09-15/train.log b/outputs/2025-12-08/09-09-15/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/10-11-41/.hydra/config.yaml b/outputs/2025-12-08/10-11-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fd8d1a9d03da83a406db364257a5251b74c406f9
--- /dev/null
+++ b/outputs/2025-12-08/10-11-41/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/10-11-41/.hydra/hydra.yaml b/outputs/2025-12-08/10-11-41/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dabce059a5a015fdfd91797f6eed425b2d1b0434
--- /dev/null
+++ b/outputs/2025-12-08/10-11-41/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=rubikscube-2
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-2,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/10-11-41
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/10-11-41/.hydra/overrides.yaml b/outputs/2025-12-08/10-11-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5f27a55696956916072c9ad292081b06707d7ba6
--- /dev/null
+++ b/outputs/2025-12-08/10-11-41/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=rubikscube-2
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/10-11-41/train.log b/outputs/2025-12-08/10-11-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/10-37-22/.hydra/config.yaml b/outputs/2025-12-08/10-37-22/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..baed48fae0b57062d1d3869e4c408d0031fe0d78
--- /dev/null
+++ b/outputs/2025-12-08/10-37-22/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: frozen_lake-ppo-slippery-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/10-37-22/.hydra/hydra.yaml b/outputs/2025-12-08/10-37-22/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..213c088b8648f88b426ed1a0b355c2c341368465
--- /dev/null
+++ b/outputs/2025-12-08/10-37-22/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=frozen_lake-ppo-slippery-starpos,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _3_frozen_lake
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/10-37-22
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/10-37-22/.hydra/overrides.yaml b/outputs/2025-12-08/10-37-22/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..add3906055cb2e1f83d5e1dda9c158870205dfc3
--- /dev/null
+++ b/outputs/2025-12-08/10-37-22/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/10-37-22/train.log b/outputs/2025-12-08/10-37-22/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/10-57-24/.hydra/config.yaml b/outputs/2025-12-08/10-57-24/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..baed48fae0b57062d1d3869e4c408d0031fe0d78
--- /dev/null
+++ b/outputs/2025-12-08/10-57-24/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: frozen_lake-ppo-slippery-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/10-57-24/.hydra/hydra.yaml b/outputs/2025-12-08/10-57-24/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c50a72fc3fca6c03f8820cc1941841b919a60cc1
--- /dev/null
+++ b/outputs/2025-12-08/10-57-24/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=frozen_lake-ppo-slippery-starpos,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _3_frozen_lake
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/10-57-24
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/10-57-24/.hydra/overrides.yaml b/outputs/2025-12-08/10-57-24/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..add3906055cb2e1f83d5e1dda9c158870205dfc3
--- /dev/null
+++ b/outputs/2025-12-08/10-57-24/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/10-57-24/train.log b/outputs/2025-12-08/10-57-24/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/11-05-03/.hydra/config.yaml b/outputs/2025-12-08/11-05-03/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..327bfe03f6012ce68b61694e386205658505d936
--- /dev/null
+++ b/outputs/2025-12-08/11-05-03/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.3
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: frozen_lake-ppo-slippery-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/11-05-03/.hydra/hydra.yaml b/outputs/2025-12-08/11-05-03/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6d2f005eb5646ce125a78f22167008fb38314e2f
--- /dev/null
+++ b/outputs/2025-12-08/11-05-03/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=frozen_lake-ppo-slippery-starpos,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _3_frozen_lake
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/11-05-03
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/11-05-03/.hydra/overrides.yaml b/outputs/2025-12-08/11-05-03/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..add3906055cb2e1f83d5e1dda9c158870205dfc3
--- /dev/null
+++ b/outputs/2025-12-08/11-05-03/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/11-05-03/train.log b/outputs/2025-12-08/11-05-03/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/11-09-45/.hydra/config.yaml b/outputs/2025-12-08/11-09-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..baed48fae0b57062d1d3869e4c408d0031fe0d78
--- /dev/null
+++ b/outputs/2025-12-08/11-09-45/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: frozen_lake-ppo-slippery-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/11-09-45/.hydra/hydra.yaml b/outputs/2025-12-08/11-09-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7c0de031a32822fc96693fadb631d64d8453aaf5
--- /dev/null
+++ b/outputs/2025-12-08/11-09-45/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=frozen_lake-ppo-slippery-starpos,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _3_frozen_lake
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/11-09-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/11-09-45/.hydra/overrides.yaml b/outputs/2025-12-08/11-09-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..add3906055cb2e1f83d5e1dda9c158870205dfc3
--- /dev/null
+++ b/outputs/2025-12-08/11-09-45/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=frozen_lake-ppo-slippery-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/11-09-45/train.log b/outputs/2025-12-08/11-09-45/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/11-18-06/.hydra/config.yaml b/outputs/2025-12-08/11-18-06/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8c80a97de74370f2082b5aeb96f55a3b813d48e1
--- /dev/null
+++ b/outputs/2025-12-08/11-18-06/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/11-18-06/.hydra/hydra.yaml b/outputs/2025-12-08/11-18-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e4ac8aa825f59a3a2ed4a948d738700b07a05948
--- /dev/null
+++ b/outputs/2025-12-08/11-18-06/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube-1
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1',trainer.experiment_name=rubikscube-1,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/11-18-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/11-18-06/.hydra/overrides.yaml b/outputs/2025-12-08/11-18-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ef46508cbe33c821d94f29b5952148497d65c2f2
--- /dev/null
+++ b/outputs/2025-12-08/11-18-06/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube-1
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/11-18-06/train.log b/outputs/2025-12-08/11-18-06/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/11-25-31/.hydra/config.yaml b/outputs/2025-12-08/11-25-31/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5ff7634705b9b27c680f5394302325c7ae030612
--- /dev/null
+++ b/outputs/2025-12-08/11-25-31/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/11-25-31/.hydra/hydra.yaml b/outputs/2025-12-08/11-25-31/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..61c50e718b9dc5268392de0456b34d7e2eed9c76
--- /dev/null
+++ b/outputs/2025-12-08/11-25-31/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2,3'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=rubikscube-2
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='2,3',trainer.experiment_name=rubikscube-2,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/11-25-31
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/11-25-31/.hydra/overrides.yaml b/outputs/2025-12-08/11-25-31/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7060e502cb77f3ae428c6d0ac1a3cbf9c0a74d50
--- /dev/null
+++ b/outputs/2025-12-08/11-25-31/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='2,3'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=rubikscube-2
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/11-25-31/train.log b/outputs/2025-12-08/11-25-31/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/11-31-28/.hydra/config.yaml b/outputs/2025-12-08/11-31-28/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d04efa66a3174f2521f8c5abbe5e5541f040f0eb
--- /dev/null
+++ b/outputs/2025-12-08/11-31-28/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/11-31-28/.hydra/hydra.yaml b/outputs/2025-12-08/11-31-28/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..89b866631f68e1255f0e0b0a89f36ecdfdd4be24
--- /dev/null
+++ b/outputs/2025-12-08/11-31-28/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=rubikscube-2
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5',trainer.experiment_name=rubikscube-2,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/11-31-28
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/11-31-28/.hydra/overrides.yaml b/outputs/2025-12-08/11-31-28/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d1f7332ebb0e91fb6ddf9fe72174d926eacccbe3
--- /dev/null
+++ b/outputs/2025-12-08/11-31-28/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=rubikscube-2
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/11-31-28/train.log b/outputs/2025-12-08/11-31-28/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/14-02-45/.hydra/config.yaml b/outputs/2025-12-08/14-02-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8d2749edc1065c8d09eb1003ea5360722a9b1295
--- /dev/null
+++ b/outputs/2025-12-08/14-02-45/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-1-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/14-02-45/.hydra/hydra.yaml b/outputs/2025-12-08/14-02-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..379f024f7397784c5015c30e82665b76037c8b33
--- /dev/null
+++ b/outputs/2025-12-08/14-02-45/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=rubikscube-1-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5',trainer.experiment_name=rubikscube-1-starpos,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/14-02-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/14-02-45/.hydra/overrides.yaml b/outputs/2025-12-08/14-02-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..02f057bda792f2c564372ba621b9b8aabe762c48
--- /dev/null
+++ b/outputs/2025-12-08/14-02-45/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=rubikscube-1-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/14-02-45/train.log b/outputs/2025-12-08/14-02-45/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/14-51-07/.hydra/config.yaml b/outputs/2025-12-08/14-51-07/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ac3e1973e17361859f89d8eb0210aefa08a2f8c8
--- /dev/null
+++ b/outputs/2025-12-08/14-51-07/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-3
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/14-51-07/.hydra/hydra.yaml b/outputs/2025-12-08/14-51-07/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..02afb3dc85f13f05911b32223813ce5b04858f3f
--- /dev/null
+++ b/outputs/2025-12-08/14-51-07/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube-3
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='6,7',trainer.experiment_name=rubikscube-3,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/14-51-07
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/14-51-07/.hydra/overrides.yaml b/outputs/2025-12-08/14-51-07/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..75ef159ad199f9ed86a58c30e9f6996550b2374d
--- /dev/null
+++ b/outputs/2025-12-08/14-51-07/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube-3
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/14-51-07/train.log b/outputs/2025-12-08/14-51-07/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/16-44-17/.hydra/config.yaml b/outputs/2025-12-08/16-44-17/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ac3e1973e17361859f89d8eb0210aefa08a2f8c8
--- /dev/null
+++ b/outputs/2025-12-08/16-44-17/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 1
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-3
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/16-44-17/.hydra/hydra.yaml b/outputs/2025-12-08/16-44-17/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..07d926b05798a1242b48d9fff8222b0829a30a26
--- /dev/null
+++ b/outputs/2025-12-08/16-44-17/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=rubikscube-3
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.001
+    - actor_rollout_ref.actor.kl_loss_coef=0.001
+    - actor_rollout_ref.actor.clip_ratio_high=0.2
+    - actor_rollout_ref.rollout.rollout_filter_ratio=1
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.2,actor_rollout_ref.actor.kl_loss_coef=0.001,actor_rollout_ref.rollout.rollout_filter_ratio=1,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.001,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='6,7',trainer.experiment_name=rubikscube-3,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/16-44-17
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/16-44-17/.hydra/overrides.yaml b/outputs/2025-12-08/16-44-17/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..75ef159ad199f9ed86a58c30e9f6996550b2374d
--- /dev/null
+++ b/outputs/2025-12-08/16-44-17/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=rubikscube-3
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.001
+- actor_rollout_ref.actor.kl_loss_coef=0.001
+- actor_rollout_ref.actor.clip_ratio_high=0.2
+- actor_rollout_ref.rollout.rollout_filter_ratio=1
diff --git a/outputs/2025-12-08/16-44-17/train.log b/outputs/2025-12-08/16-44-17/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/16-54-39/.hydra/config.yaml b/outputs/2025-12-08/16-54-39/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..acfd529b6bbf31cbf3a89b726176c3f0a388320a
--- /dev/null
+++ b/outputs/2025-12-08/16-54-39/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/16-54-39/.hydra/hydra.yaml b/outputs/2025-12-08/16-54-39/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e731254d79894d0f7862173b7e8ac36bfe0eb9de
--- /dev/null
+++ b/outputs/2025-12-08/16-54-39/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=rubikscube-1-starpos,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/16-54-39
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/16-54-39/.hydra/overrides.yaml b/outputs/2025-12-08/16-54-39/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..71ee7689b21e6bfd0114f266aadcd61b72d8231f
--- /dev/null
+++ b/outputs/2025-12-08/16-54-39/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- trainer.experiment_name=rubikscube-1-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/16-54-39/train.log b/outputs/2025-12-08/16-54-39/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/17-11-46/.hydra/config.yaml b/outputs/2025-12-08/17-11-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a3ef1c791b4414c41e4fbeba4117faa7da4d4e8d
--- /dev/null
+++ b/outputs/2025-12-08/17-11-46/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '3'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/17-11-46/.hydra/hydra.yaml b/outputs/2025-12-08/17-11-46/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e8fb474e6c840bb58b2c7b584935c1a0e0ef3dac
--- /dev/null
+++ b/outputs/2025-12-08/17-11-46/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='3'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='3',trainer.experiment_name=rubikscube-3-starpos,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/17-11-46
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/17-11-46/.hydra/overrides.yaml b/outputs/2025-12-08/17-11-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cdd98455bea0328b0b24fad64db33a7f9cae25b5
--- /dev/null
+++ b/outputs/2025-12-08/17-11-46/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='3'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/17-11-46/train.log b/outputs/2025-12-08/17-11-46/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/20-11-22/.hydra/config.yaml b/outputs/2025-12-08/20-11-22/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ccc80239811405e71df4162504534bbbc6b128af
--- /dev/null
+++ b/outputs/2025-12-08/20-11-22/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-3-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_3
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/20-11-22/.hydra/hydra.yaml b/outputs/2025-12-08/20-11-22/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8294573539fd0933f480c3030441689ce9779977
--- /dev/null
+++ b/outputs/2025-12-08/20-11-22/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_3
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_3,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=rubikscube-3-starpos-frommlp,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/20-11-22
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/20-11-22/.hydra/overrides.yaml b/outputs/2025-12-08/20-11-22/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0958649014230318447a355e7abdbe3c46b7667d
--- /dev/null
+++ b/outputs/2025-12-08/20-11-22/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_3
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/20-11-22/train.log b/outputs/2025-12-08/20-11-22/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/20-37-35/.hydra/config.yaml b/outputs/2025-12-08/20-37-35/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b56b2b2405d57f0b7adc39309bddf0b611b231be
--- /dev/null
+++ b/outputs/2025-12-08/20-37-35/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '2'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/20-37-35/.hydra/hydra.yaml b/outputs/2025-12-08/20-37-35/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8401addcc67e4da90f71b02828315245a800952a
--- /dev/null
+++ b/outputs/2025-12-08/20-37-35/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1,system.CUDA_VISIBLE_DEVICES='2',trainer.experiment_name=rubikscube-1-starpos-frommlp,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/20-37-35
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/20-37-35/.hydra/overrides.yaml b/outputs/2025-12-08/20-37-35/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5457aacefb6e57638ec9001680cb5c9f1e3c5ba6
--- /dev/null
+++ b/outputs/2025-12-08/20-37-35/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='2'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- trainer.experiment_name=rubikscube-1-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/20-37-35/train.log b/outputs/2025-12-08/20-37-35/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/20-48-28/.hydra/config.yaml b/outputs/2025-12-08/20-48-28/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2e3422ccfa4d7a8096b5e62f4405d92bcf2ea7c5
--- /dev/null
+++ b/outputs/2025-12-08/20-48-28/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/20-48-28/.hydra/hydra.yaml b/outputs/2025-12-08/20-48-28/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..04ee313b22061665b6027e2dc148eb2081ef3d5e
--- /dev/null
+++ b/outputs/2025-12-08/20-48-28/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2,3'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='2,3',trainer.experiment_name=rubikscube-2-starpos,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/20-48-28
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/20-48-28/.hydra/overrides.yaml b/outputs/2025-12-08/20-48-28/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fee48d983c8c66e5a758f83b51d81a1368a40c59
--- /dev/null
+++ b/outputs/2025-12-08/20-48-28/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='2,3'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/20-48-28/train.log b/outputs/2025-12-08/20-48-28/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/22-09-38/.hydra/config.yaml b/outputs/2025-12-08/22-09-38/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..25da20758e28004f160893e57604f09607b16c32
--- /dev/null
+++ b/outputs/2025-12-08/22-09-38/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/22-09-38/.hydra/hydra.yaml b/outputs/2025-12-08/22-09-38/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b5329e3df427748e959f57e6f8bcb94d3c249c41
--- /dev/null
+++ b/outputs/2025-12-08/22-09-38/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1',trainer.experiment_name=rubikscube-3-starpos,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/22-09-38
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/22-09-38/.hydra/overrides.yaml b/outputs/2025-12-08/22-09-38/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cda87fa4b1152d583c8c6b68b5e393cec67bd0e6
--- /dev/null
+++ b/outputs/2025-12-08/22-09-38/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/22-09-38/train.log b/outputs/2025-12-08/22-09-38/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/22-18-27/.hydra/config.yaml b/outputs/2025-12-08/22-18-27/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..81a426671ab4ecaf88223db8e7575d6d0af15107
--- /dev/null
+++ b/outputs/2025-12-08/22-18-27/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-1-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_1
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/22-18-27/.hydra/hydra.yaml b/outputs/2025-12-08/22-18-27/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..522ef0efa1e0ec90f5a3fa5573366359bacffd6f
--- /dev/null
+++ b/outputs/2025-12-08/22-18-27/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_1
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_1,system.CUDA_VISIBLE_DEVICES='6,7',trainer.experiment_name=rubikscube-1-starpos-frommlp,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/22-18-27
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/22-18-27/.hydra/overrides.yaml b/outputs/2025-12-08/22-18-27/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3dc6595938af2849185d5d164b1f970486761882
--- /dev/null
+++ b/outputs/2025-12-08/22-18-27/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_1
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- trainer.experiment_name=rubikscube-1-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/22-18-27/train.log b/outputs/2025-12-08/22-18-27/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-08/22-36-50/.hydra/config.yaml b/outputs/2025-12-08/22-36-50/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..99d568fecaaa8ae47450601f06ef7a1ce02a4511
--- /dev/null
+++ b/outputs/2025-12-08/22-36-50/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_2
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-08/22-36-50/.hydra/hydra.yaml b/outputs/2025-12-08/22-36-50/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d9126058afe0adc58041239bdb6573644b6abab5
--- /dev/null
+++ b/outputs/2025-12-08/22-36-50/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_2
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_2,system.CUDA_VISIBLE_DEVICES='4,5',trainer.experiment_name=rubikscube-2-starpos-frommlp,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-08/22-36-50
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-08/22-36-50/.hydra/overrides.yaml b/outputs/2025-12-08/22-36-50/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f2ff72424b793b2e6e4d35e528aeae81d261411a
--- /dev/null
+++ b/outputs/2025-12-08/22-36-50/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='4,5'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_2
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-08/22-36-50/train.log b/outputs/2025-12-08/22-36-50/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/00-00-46/.hydra/config.yaml b/outputs/2025-12-09/00-00-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8fa01f3e92b951829ac93835a55f073b1bf709ce
--- /dev/null
+++ b/outputs/2025-12-09/00-00-46/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/00-00-46/.hydra/hydra.yaml b/outputs/2025-12-09/00-00-46/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f69e5e9abc0c48eba800f8c38ad2093ee16d81c6
--- /dev/null
+++ b/outputs/2025-12-09/00-00-46/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3,system.CUDA_VISIBLE_DEVICES='4,5',trainer.experiment_name=rubikscube-3-starpos-frommlp,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/00-00-46
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/00-00-46/.hydra/overrides.yaml b/outputs/2025-12-09/00-00-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a1c55dde178a42c2c9d592bbca9fd81edd4664a7
--- /dev/null
+++ b/outputs/2025-12-09/00-00-46/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='4,5'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/00-00-46/train.log b/outputs/2025-12-09/00-00-46/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/01-24-32/.hydra/config.yaml b/outputs/2025-12-09/01-24-32/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8fa01f3e92b951829ac93835a55f073b1bf709ce
--- /dev/null
+++ b/outputs/2025-12-09/01-24-32/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/01-24-32/.hydra/hydra.yaml b/outputs/2025-12-09/01-24-32/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e3ba9a4c02362b170fa2941667ac9e3669c79c89
--- /dev/null
+++ b/outputs/2025-12-09/01-24-32/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3,system.CUDA_VISIBLE_DEVICES='4,5',trainer.experiment_name=rubikscube-3-starpos-frommlp,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/01-24-32
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/01-24-32/.hydra/overrides.yaml b/outputs/2025-12-09/01-24-32/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a1c55dde178a42c2c9d592bbca9fd81edd4664a7
--- /dev/null
+++ b/outputs/2025-12-09/01-24-32/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='4,5'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/01-24-32/train.log b/outputs/2025-12-09/01-24-32/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/09-09-41/.hydra/config.yaml b/outputs/2025-12-09/09-09-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cc56c7cfb050effcb922ef8ee16dcc6dc2b4aa8f
--- /dev/null
+++ b/outputs/2025-12-09/09-09-41/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/09-09-41/.hydra/hydra.yaml b/outputs/2025-12-09/09-09-41/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..df7b49b9dddcf89f81568362ab1c5e14530e49e2
--- /dev/null
+++ b/outputs/2025-12-09/09-09-41/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=rubikscube-1-starpos-frommlp,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/09-09-41
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/09-09-41/.hydra/overrides.yaml b/outputs/2025-12-09/09-09-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b0867a9360f34cab5e4b47261f74b7855ff0733e
--- /dev/null
+++ b/outputs/2025-12-09/09-09-41/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_1
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- trainer.experiment_name=rubikscube-1-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/09-09-41/train.log b/outputs/2025-12-09/09-09-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/09-09-42/.hydra/config.yaml b/outputs/2025-12-09/09-09-42/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dbb80f36c0f4353bc5907619bafc77efc2d27c4d
--- /dev/null
+++ b/outputs/2025-12-09/09-09-42/.hydra/config.yaml
@@ -0,0 +1,806 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '1'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_2
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+ll
diff --git a/outputs/2025-12-09/09-09-42/.hydra/hydra.yaml b/outputs/2025-12-09/09-09-42/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1f8affc9b3c2a4b9fc7ebf23a6f085e590999ed4
--- /dev/null
+++ b/outputs/2025-12-09/09-09-42/.hydra/hydra.yaml
@@ -0,0 +1,181 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='1'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_2
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_2,system.CUDA_VISIBLE_DEVICES='1',trainer.experiment_name=rubikscube-2-starpos-frommlp,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/09-09-42
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
+se: false
diff --git a/outputs/2025-12-09/09-09-42/.hydra/overrides.yaml b/outputs/2025-12-09/09-09-42/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..02d4576c671846853687660ebaa0783c81dbdfba
--- /dev/null
+++ b/outputs/2025-12-09/09-09-42/.hydra/overrides.yaml
@@ -0,0 +1,12 @@
+- system.CUDA_VISIBLE_DEVICES='1'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_2
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+0.25
diff --git a/outputs/2025-12-09/09-09-42/train.log b/outputs/2025-12-09/09-09-42/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/10-03-00/.hydra/config.yaml b/outputs/2025-12-09/10-03-00/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2cb054673be247362134212467ebdf0f01f3d127
--- /dev/null
+++ b/outputs/2025-12-09/10-03-00/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 3,4
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/10-03-00/.hydra/hydra.yaml b/outputs/2025-12-09/10-03-00/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..32614602ae6f548b72c3828406ef8561ba6dfb6a
--- /dev/null
+++ b/outputs/2025-12-09/10-03-00/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='3,4'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3,system.CUDA_VISIBLE_DEVICES='3,4',trainer.experiment_name=rubikscube-3-starpos-frommlp,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/10-03-00
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/10-03-00/.hydra/overrides.yaml b/outputs/2025-12-09/10-03-00/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d22f4feaad276b0fef9bd429b702fcc4c2fc8e95
--- /dev/null
+++ b/outputs/2025-12-09/10-03-00/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='3,4'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube_3
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/10-03-00/train.log b/outputs/2025-12-09/10-03-00/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/10-25-05/.hydra/config.yaml b/outputs/2025-12-09/10-25-05/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..be986dfc9f05f7babe234fa36b33fdc07647843b
--- /dev/null
+++ b/outputs/2025-12-09/10-25-05/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: game_2048
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 2
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 5,6
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/10-25-05/.hydra/hydra.yaml b/outputs/2025-12-09/10-25-05/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e5d049f3c79e1e53e43ede9b9f73f08729e4d130
--- /dev/null
+++ b/outputs/2025-12-09/10-25-05/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='5,6'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=game_2048
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='5,6',trainer.experiment_name=game_2048,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/10-25-05
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/10-25-05/.hydra/overrides.yaml b/outputs/2025-12-09/10-25-05/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dc802c13401b9fb39a7ed128857f10f60daeaa11
--- /dev/null
+++ b/outputs/2025-12-09/10-25-05/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='5,6'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=2
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=game_2048
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/10-25-05/train.log b/outputs/2025-12-09/10-25-05/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/11-17-41/.hydra/config.yaml b/outputs/2025-12-09/11-17-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d5b21f7bece1fa6c5f05b7fbb2820933bdc8c98e
--- /dev/null
+++ b/outputs/2025-12-09/11-17-41/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-1-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 2,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_1
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/11-17-41/.hydra/hydra.yaml b/outputs/2025-12-09/11-17-41/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..26b576be4b9e11b01a6b9b44b15bb41a540d9f34
--- /dev/null
+++ b/outputs/2025-12-09/11-17-41/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='2,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_1
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_1,system.CUDA_VISIBLE_DEVICES='2,5,6,7',trainer.experiment_name=rubikscube-1-starpos-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/11-17-41
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/11-17-41/.hydra/overrides.yaml b/outputs/2025-12-09/11-17-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2dba320015f26c3f2aa5377714fab2e9f52dd7ba
--- /dev/null
+++ b/outputs/2025-12-09/11-17-41/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='2,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_1
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- trainer.experiment_name=rubikscube-1-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/11-17-41/train.log b/outputs/2025-12-09/11-17-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/12-16-01/.hydra/config.yaml b/outputs/2025-12-09/12-16-01/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4864d4afa5407c6d6527487807c73e84812d09b3
--- /dev/null
+++ b/outputs/2025-12-09/12-16-01/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/12-16-01/.hydra/hydra.yaml b/outputs/2025-12-09/12-16-01/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fd1a0714340893b7bcd81807ebbbcaa258253470
--- /dev/null
+++ b/outputs/2025-12-09/12-16-01/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=rubikscube-2-starpos-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/12-16-01
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/12-16-01/.hydra/overrides.yaml b/outputs/2025-12-09/12-16-01/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..62c4030aaf4ee15330e72cc27e1e250f3e9a88a4
--- /dev/null
+++ b/outputs/2025-12-09/12-16-01/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/12-16-01/train.log b/outputs/2025-12-09/12-16-01/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/12-17-06/.hydra/config.yaml b/outputs/2025-12-09/12-17-06/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b9f085d8865c0faa4276602ea391a6a30fcb65ff
--- /dev/null
+++ b/outputs/2025-12-09/12-17-06/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/12-17-06/.hydra/hydra.yaml b/outputs/2025-12-09/12-17-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..30f7bac771bad505d961c62d4a92c748a45faf16
--- /dev/null
+++ b/outputs/2025-12-09/12-17-06/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/12-17-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/12-17-06/.hydra/overrides.yaml b/outputs/2025-12-09/12-17-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d304c2dd3f1306ba0c3d3cca6ad6ecc392176862
--- /dev/null
+++ b/outputs/2025-12-09/12-17-06/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/12-17-06/train.log b/outputs/2025-12-09/12-17-06/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/14-14-46/.hydra/config.yaml b/outputs/2025-12-09/14-14-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..93fa77cb58002f691f429c2de0b16e8e6ac9523b
--- /dev/null
+++ b/outputs/2025-12-09/14-14-46/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/14-14-46/.hydra/hydra.yaml b/outputs/2025-12-09/14-14-46/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4760aaf4e95303ca3cba2c1330bfbe58a501607e
--- /dev/null
+++ b/outputs/2025-12-09/14-14-46/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/14-14-46
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/14-14-46/.hydra/overrides.yaml b/outputs/2025-12-09/14-14-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3324ec0f5c0c74d30c408aa82c96ab2ed7baae9c
--- /dev/null
+++ b/outputs/2025-12-09/14-14-46/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_2
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/14-14-46/train.log b/outputs/2025-12-09/14-14-46/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/14-37-30/.hydra/config.yaml b/outputs/2025-12-09/14-37-30/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b9f085d8865c0faa4276602ea391a6a30fcb65ff
--- /dev/null
+++ b/outputs/2025-12-09/14-37-30/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/14-37-30/.hydra/hydra.yaml b/outputs/2025-12-09/14-37-30/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8480fbaee259ed8981e5ece6be6dd7c94e1b1597
--- /dev/null
+++ b/outputs/2025-12-09/14-37-30/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/14-37-30
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/14-37-30/.hydra/overrides.yaml b/outputs/2025-12-09/14-37-30/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d304c2dd3f1306ba0c3d3cca6ad6ecc392176862
--- /dev/null
+++ b/outputs/2025-12-09/14-37-30/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube_3
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/14-37-30/train.log b/outputs/2025-12-09/14-37-30/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/15-06-27/.hydra/config.yaml b/outputs/2025-12-09/15-06-27/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f75f5a2ff3609cdfbb3bb7d96f829c484ec5cd2e
--- /dev/null
+++ b/outputs/2025-12-09/15-06-27/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-09/15-06-27/.hydra/hydra.yaml b/outputs/2025-12-09/15-06-27/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9bc8308e92a725c353870dba88b50d4e335e6310
--- /dev/null
+++ b/outputs/2025-12-09/15-06-27/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=rubikscube-3-starpos,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-09/15-06-27
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-09/15-06-27/.hydra/overrides.yaml b/outputs/2025-12-09/15-06-27/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f544fc730abe122ada71d403ead4a681801735e0
--- /dev/null
+++ b/outputs/2025-12-09/15-06-27/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/15-06-27/train.log b/outputs/2025-12-09/15-06-27/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-09/17-02-50/.hydra/overrides.yaml b/outputs/2025-12-09/17-02-50/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..75ff1c16597fcbbed064470fb3ed55fa443201ff
--- /dev/null
+++ b/outputs/2025-12-09/17-02-50/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='3,4,5,6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=game_2048
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-09/17-02-50/train.log b/outputs/2025-12-09/17-02-50/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/00-49-58/.hydra/config.yaml b/outputs/2025-12-10/00-49-58/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2d0e98ac61f33e6582536e83bc1efeba66d6e894
--- /dev/null
+++ b/outputs/2025-12-10/00-49-58/.hydra/config.yaml
@@ -0,0 +1,806 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+ll
diff --git a/outputs/2025-12-10/00-49-58/.hydra/overrides.yaml b/outputs/2025-12-10/00-49-58/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d57ff3564fadba7158ed15d62f8d20eb18274f64
--- /dev/null
+++ b/outputs/2025-12-10/00-49-58/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=game_2048
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-10/00-49-58/train.log b/outputs/2025-12-10/00-49-58/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/11-03-21/.hydra/config.yaml b/outputs/2025-12-10/11-03-21/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1d2a079e131d547d7ca9ab8885a2a015353c3907
--- /dev/null
+++ b/outputs/2025-12-10/11-03-21/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-10/11-03-21/.hydra/hydra.yaml b/outputs/2025-12-10/11-03-21/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6a1dfe272be5bf0d906dd5689087c17d1606f6f5
--- /dev/null
+++ b/outputs/2025-12-10/11-03-21/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=game_2048
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=game_2048,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-10/11-03-21
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-10/11-03-21/.hydra/overrides.yaml b/outputs/2025-12-10/11-03-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f5ead8086f5552dc80b6b98c1f0e7102535d446
--- /dev/null
+++ b/outputs/2025-12-10/11-03-21/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=game_2048
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-10/11-03-21/train.log b/outputs/2025-12-10/11-03-21/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/11-10-41/.hydra/config.yaml b/outputs/2025-12-10/11-10-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1d2a079e131d547d7ca9ab8885a2a015353c3907
--- /dev/null
+++ b/outputs/2025-12-10/11-10-41/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-10/11-10-41/.hydra/overrides.yaml b/outputs/2025-12-10/11-10-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f5ead8086f5552dc80b6b98c1f0e7102535d446
--- /dev/null
+++ b/outputs/2025-12-10/11-10-41/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=game_2048
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-10/11-10-41/train.log b/outputs/2025-12-10/11-10-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/11-38-06/.hydra/hydra.yaml b/outputs/2025-12-10/11-38-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7c7780301fb4b32ba85f902fcacbd4edcf86c2e1
--- /dev/null
+++ b/outputs/2025-12-10/11-38-06/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=game_2048
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=game_2048,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-10/11-38-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-10/11-38-06/train.log b/outputs/2025-12-10/11-38-06/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/11-39-52/.hydra/overrides.yaml b/outputs/2025-12-10/11-39-52/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f5ead8086f5552dc80b6b98c1f0e7102535d446
--- /dev/null
+++ b/outputs/2025-12-10/11-39-52/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=game_2048
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-10/11-39-52/train.log b/outputs/2025-12-10/11-39-52/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/11-44-45/.hydra/config.yaml b/outputs/2025-12-10/11-44-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1d2a079e131d547d7ca9ab8885a2a015353c3907
--- /dev/null
+++ b/outputs/2025-12-10/11-44-45/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-10/11-44-45/.hydra/overrides.yaml b/outputs/2025-12-10/11-44-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f5ead8086f5552dc80b6b98c1f0e7102535d446
--- /dev/null
+++ b/outputs/2025-12-10/11-44-45/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=game_2048
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-10/11-44-45/train.log b/outputs/2025-12-10/11-44-45/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/11-51-31/.hydra/config.yaml b/outputs/2025-12-10/11-51-31/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1d2a079e131d547d7ca9ab8885a2a015353c3907
--- /dev/null
+++ b/outputs/2025-12-10/11-51-31/.hydra/config.yaml
@@ -0,0 +1,805 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-10/11-58-26/train.log b/outputs/2025-12-10/11-58-26/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/12-37-48/train.log b/outputs/2025-12-10/12-37-48/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-10/19-44-44/train.log b/outputs/2025-12-10/19-44-44/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-11/15-16-21/train.log b/outputs/2025-12-11/15-16-21/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-11/15-25-10/.hydra/hydra.yaml b/outputs/2025-12-11/15-25-10/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c372bae2039e791962f3b3f95d5e2ed36dc11a65
--- /dev/null
+++ b/outputs/2025-12-11/15-25-10/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=sudoku-4x4,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-11/15-25-10
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-11/15-43-08/.hydra/config.yaml b/outputs/2025-12-11/15-43-08/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2542d70a880823a5f96dba1cb6e1ab918c222ba1
--- /dev/null
+++ b/outputs/2025-12-11/15-43-08/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-11/22-03-13/.hydra/hydra.yaml b/outputs/2025-12-11/22-03-13/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..95a0ad201b7430adf396e09efc73f959a9ff74d8
--- /dev/null
+++ b/outputs/2025-12-11/22-03-13/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='1,2'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4
+    - trainer.experiment_name=sudoku-4x4-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4,system.CUDA_VISIBLE_DEVICES='1,2',trainer.experiment_name=sudoku-4x4-frommlp,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-11/22-03-13
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-11/22-12-11/.hydra/config.yaml b/outputs/2025-12-11/22-12-11/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0043531ae0646b4aa17c24645d5710a15afefdc7
--- /dev/null
+++ b/outputs/2025-12-11/22-12-11/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.6
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4-frommlp
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 3,4,5,6
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-11/22-12-11/.hydra/overrides.yaml b/outputs/2025-12-11/22-12-11/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..255bac5e33b7c920f8939e3e94aebece10d97fdc
--- /dev/null
+++ b/outputs/2025-12-11/22-12-11/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='3,4,5,6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4
+- trainer.experiment_name=sudoku-4x4-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-11/22-29-11/.hydra/hydra.yaml b/outputs/2025-12-11/22-29-11/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..985790ef638d2f70e567d1de044cf8fea9fc63cb
--- /dev/null
+++ b/outputs/2025-12-11/22-29-11/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='1,2'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=2
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4
+    - trainer.experiment_name=sudoku-4x4-frommlp
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4,system.CUDA_VISIBLE_DEVICES='1,2',trainer.experiment_name=sudoku-4x4-frommlp,trainer.n_gpus_per_node=2,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-11/22-29-11
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-11/23-07-38/.hydra/overrides.yaml b/outputs/2025-12-11/23-07-38/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..255bac5e33b7c920f8939e3e94aebece10d97fdc
--- /dev/null
+++ b/outputs/2025-12-11/23-07-38/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='3,4,5,6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4
+- trainer.experiment_name=sudoku-4x4-frommlp
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/06-00-18/.hydra/overrides.yaml b/outputs/2025-12-12/06-00-18/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a51c0aa037effc6ee4b17e2bed26a75d81219020
--- /dev/null
+++ b/outputs/2025-12-12/06-00-18/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=sudoku-4x4
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/08-29-04/train.log b/outputs/2025-12-12/08-29-04/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/08-38-31/.hydra/config.yaml b/outputs/2025-12-12/08-38-31/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1f229888570e4e7419a0b453d11af8bcb2023c42
--- /dev/null
+++ b/outputs/2025-12-12/08-38-31/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048-frommlp-windows5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/08-38-31/.hydra/hydra.yaml b/outputs/2025-12-12/08-38-31/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..66656352e646c35916e38614b99fc2406b7719b0
--- /dev/null
+++ b/outputs/2025-12-12/08-38-31/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+    - trainer.experiment_name=game_2048-frommlp-windows5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=game_2048-frommlp-windows5,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/08-38-31
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/08-38-31/.hydra/overrides.yaml b/outputs/2025-12-12/08-38-31/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a7fadbdf30c0dad909ddda674317b312e2a8ac07
--- /dev/null
+++ b/outputs/2025-12-12/08-38-31/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+- trainer.experiment_name=game_2048-frommlp-windows5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/08-38-31/train.log b/outputs/2025-12-12/08-38-31/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/09-54-20/.hydra/config.yaml b/outputs/2025-12-12/09-54-20/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ea20c93d6e19509ff89b344868fd303c31245b4a
--- /dev/null
+++ b/outputs/2025-12-12/09-54-20/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/09-54-20/.hydra/hydra.yaml b/outputs/2025-12-12/09-54-20/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8589a4c1622c0c8da344cbf0077ce143ad0fae9b
--- /dev/null
+++ b/outputs/2025-12-12/09-54-20/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=sudoku-4x4
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=sudoku-4x4,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/09-54-20
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/09-54-20/train.log b/outputs/2025-12-12/09-54-20/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/10-51-01/.hydra/config.yaml b/outputs/2025-12-12/10-51-01/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..837c7f750d41423e80d1e455fdc502e4108d23af
--- /dev/null
+++ b/outputs/2025-12-12/10-51-01/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.3
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/10-51-01/.hydra/hydra.yaml b/outputs/2025-12-12/10-51-01/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..70584292ce631599d9a1bb30722f41cd4182d14b
--- /dev/null
+++ b/outputs/2025-12-12/10-51-01/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=sudoku-4x4
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=sudoku-4x4,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/10-51-01
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/10-51-01/.hydra/overrides.yaml b/outputs/2025-12-12/10-51-01/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d3d382ca0f650601cc97b670ab6d72fdd2eb0b63
--- /dev/null
+++ b/outputs/2025-12-12/10-51-01/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=sudoku-4x4
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/10-58-45/.hydra/config.yaml b/outputs/2025-12-12/10-58-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..afcde3ca315be979ea813ce4dd8f37985d3e8419
--- /dev/null
+++ b/outputs/2025-12-12/10-58-45/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/10-58-45/.hydra/hydra.yaml b/outputs/2025-12-12/10-58-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..81a47030b4ecf4b9f16f051896a2137f82eba2f0
--- /dev/null
+++ b/outputs/2025-12-12/10-58-45/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=sudoku-4x4
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='4,5,6,7',trainer.experiment_name=sudoku-4x4,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/10-58-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/11-49-29/.hydra/config.yaml b/outputs/2025-12-12/11-49-29/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9d4d36a046884cf77f74b3729b7d55eb5d457c9c
--- /dev/null
+++ b/outputs/2025-12-12/11-49-29/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.6
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/11-49-29/.hydra/hydra.yaml b/outputs/2025-12-12/11-49-29/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b62cb6c6efa8aa04792c9eda90e28de2698e7f2a
--- /dev/null
+++ b/outputs/2025-12-12/11-49-29/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=sudoku-4x4,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/11-49-29
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/11-49-29/.hydra/overrides.yaml b/outputs/2025-12-12/11-49-29/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6cd891ae012b4c6ea5832828ebd944f31c2f3054
--- /dev/null
+++ b/outputs/2025-12-12/11-49-29/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=sudoku-4x4
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/11-49-29/train.log b/outputs/2025-12-12/11-49-29/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/13-55-43/.hydra/config.yaml b/outputs/2025-12-12/13-55-43/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cc65b012d3aa3922d49fa2678d07d0c4afc6f423
--- /dev/null
+++ b/outputs/2025-12-12/13-55-43/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.6
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '7'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/13-55-43/.hydra/hydra.yaml b/outputs/2025-12-12/13-55-43/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..63d9fe778e1e3d08bb102d72ba0b029433174778
--- /dev/null
+++ b/outputs/2025-12-12/13-55-43/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='7',trainer.experiment_name=sudoku-4x4,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/13-55-43
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/13-55-43/train.log b/outputs/2025-12-12/13-55-43/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/14-13-36/.hydra/config.yaml b/outputs/2025-12-12/14-13-36/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8cb327cc98a6301b6398d5ccfe0c2ecedf950f8a
--- /dev/null
+++ b/outputs/2025-12-12/14-13-36/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.6
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/14-13-36/.hydra/hydra.yaml b/outputs/2025-12-12/14-13-36/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..627ddae47e2964c7e7329bde0786d7a8c2285e2e
--- /dev/null
+++ b/outputs/2025-12-12/14-13-36/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/14-13-36
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/14-13-36/.hydra/overrides.yaml b/outputs/2025-12-12/14-13-36/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..275e5d5ab4a4c08b87e7e2c727797860ccd4fe0b
--- /dev/null
+++ b/outputs/2025-12-12/14-13-36/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/14-13-36/train.log b/outputs/2025-12-12/14-13-36/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/14-14-51/.hydra/config.yaml b/outputs/2025-12-12/14-14-51/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8cb327cc98a6301b6398d5ccfe0c2ecedf950f8a
--- /dev/null
+++ b/outputs/2025-12-12/14-14-51/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.6
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/14-14-51/.hydra/hydra.yaml b/outputs/2025-12-12/14-14-51/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2eba1d4e9d8700bf1482789e113a4e633ed16c7f
--- /dev/null
+++ b/outputs/2025-12-12/14-14-51/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/14-14-51
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/14-14-51/.hydra/overrides.yaml b/outputs/2025-12-12/14-14-51/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..275e5d5ab4a4c08b87e7e2c727797860ccd4fe0b
--- /dev/null
+++ b/outputs/2025-12-12/14-14-51/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/14-14-51/train.log b/outputs/2025-12-12/14-14-51/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/14-57-12/.hydra/config.yaml b/outputs/2025-12-12/14-57-12/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..89b462dce6d3e8581396b6175c260fb5617fbb3b
--- /dev/null
+++ b/outputs/2025-12-12/14-57-12/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.6
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/14-57-12/.hydra/hydra.yaml b/outputs/2025-12-12/14-57-12/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a254c89974d987c1663fd5aafd50b962eabf69fb
--- /dev/null
+++ b/outputs/2025-12-12/14-57-12/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/14-57-12
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/14-57-12/.hydra/overrides.yaml b/outputs/2025-12-12/14-57-12/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..29b11e22dce3e278dd2b2b0a184725a7de3e96ea
--- /dev/null
+++ b/outputs/2025-12-12/14-57-12/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/14-57-12/train.log b/outputs/2025-12-12/14-57-12/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/16-02-43/.hydra/config.yaml b/outputs/2025-12-12/16-02-43/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bbd8ecae5915b41349ff6e01d1255c4f54c75c27
--- /dev/null
+++ b/outputs/2025-12-12/16-02-43/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/16-02-43/.hydra/hydra.yaml b/outputs/2025-12-12/16-02-43/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..06ab2ee7aadb884a091c71e8ef26dcfe520610de
--- /dev/null
+++ b/outputs/2025-12-12/16-02-43/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/16-02-43
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/16-02-43/.hydra/overrides.yaml b/outputs/2025-12-12/16-02-43/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..29b11e22dce3e278dd2b2b0a184725a7de3e96ea
--- /dev/null
+++ b/outputs/2025-12-12/16-02-43/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/16-02-43/train.log b/outputs/2025-12-12/16-02-43/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/22-36-42/.hydra/config.yaml b/outputs/2025-12-12/22-36-42/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..24e7324045d7bd5e703734a9e167dbe3758d28a9
--- /dev/null
+++ b/outputs/2025-12-12/22-36-42/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/22-36-42/.hydra/hydra.yaml b/outputs/2025-12-12/22-36-42/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..78433d80c1d1898960c012467399126be9b5b4be
--- /dev/null
+++ b/outputs/2025-12-12/22-36-42/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/22-36-42
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/22-36-42/.hydra/overrides.yaml b/outputs/2025-12-12/22-36-42/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3c432e9bf3f867fa918d3986b018cb343674c59a
--- /dev/null
+++ b/outputs/2025-12-12/22-36-42/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/22-36-42/train.log b/outputs/2025-12-12/22-36-42/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/23-40-41/.hydra/config.yaml b/outputs/2025-12-12/23-40-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..11b593cdbb65a13ed85f51df9973bc9c5a4b7391
--- /dev/null
+++ b/outputs/2025-12-12/23-40-41/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/23-40-41/.hydra/hydra.yaml b/outputs/2025-12-12/23-40-41/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..70f15bba80f6257ee4364db9e50f0ce720663343
--- /dev/null
+++ b/outputs/2025-12-12/23-40-41/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/23-40-41
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/23-40-41/.hydra/overrides.yaml b/outputs/2025-12-12/23-40-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..275e5d5ab4a4c08b87e7e2c727797860ccd4fe0b
--- /dev/null
+++ b/outputs/2025-12-12/23-40-41/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/23-40-41/train.log b/outputs/2025-12-12/23-40-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/23-51-23/.hydra/config.yaml b/outputs/2025-12-12/23-51-23/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9a8af851d2f99fa9adcc95a0b08b0f15d9d0306c
--- /dev/null
+++ b/outputs/2025-12-12/23-51-23/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/23-51-23/.hydra/hydra.yaml b/outputs/2025-12-12/23-51-23/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8657854b37694eda61eebed3c81d2e13ea667e6d
--- /dev/null
+++ b/outputs/2025-12-12/23-51-23/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/23-51-23
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/23-51-23/.hydra/overrides.yaml b/outputs/2025-12-12/23-51-23/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3c432e9bf3f867fa918d3986b018cb343674c59a
--- /dev/null
+++ b/outputs/2025-12-12/23-51-23/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/23-51-23/train.log b/outputs/2025-12-12/23-51-23/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-12/23-52-03/.hydra/config.yaml b/outputs/2025-12-12/23-52-03/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9a8af851d2f99fa9adcc95a0b08b0f15d9d0306c
--- /dev/null
+++ b/outputs/2025-12-12/23-52-03/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-12/23-52-03/.hydra/hydra.yaml b/outputs/2025-12-12/23-52-03/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..eeab51f478659a57d7f4527f8851190c710bef39
--- /dev/null
+++ b/outputs/2025-12-12/23-52-03/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=4
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-12/23-52-03
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-12/23-52-03/.hydra/overrides.yaml b/outputs/2025-12-12/23-52-03/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3c432e9bf3f867fa918d3986b018cb343674c59a
--- /dev/null
+++ b/outputs/2025-12-12/23-52-03/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=4
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-12/23-52-03/train.log b/outputs/2025-12-12/23-52-03/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/00-12-53/.hydra/config.yaml b/outputs/2025-12-13/00-12-53/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..89f775a2c9dbda70005e67ffd39e7458056dae19
--- /dev/null
+++ b/outputs/2025-12-13/00-12-53/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/00-12-53/.hydra/hydra.yaml b/outputs/2025-12-13/00-12-53/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..71eb8abe2d06aa0d9100f5681e5ae148e982d7d0
--- /dev/null
+++ b/outputs/2025-12-13/00-12-53/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/00-12-53
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/00-12-53/.hydra/overrides.yaml b/outputs/2025-12-13/00-12-53/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dd0c8274e210a57b5322edf4f35630027a33dfa9
--- /dev/null
+++ b/outputs/2025-12-13/00-12-53/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/00-12-53/train.log b/outputs/2025-12-13/00-12-53/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/02-14-42/.hydra/config.yaml b/outputs/2025-12-13/02-14-42/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..786a71c323cbed30ab9a900256aa930e23e64e9b
--- /dev/null
+++ b/outputs/2025-12-13/02-14-42/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/02-14-42/.hydra/hydra.yaml b/outputs/2025-12-13/02-14-42/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5cc5c9f624105ed63303d5925e8cf20e67fbec90
--- /dev/null
+++ b/outputs/2025-12-13/02-14-42/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/02-14-42
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/02-14-42/.hydra/overrides.yaml b/outputs/2025-12-13/02-14-42/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9b0c55fce8f8c952fb4742bff85d17e07f80c09a
--- /dev/null
+++ b/outputs/2025-12-13/02-14-42/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/02-14-42/train.log b/outputs/2025-12-13/02-14-42/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/02-25-00/.hydra/config.yaml b/outputs/2025-12-13/02-25-00/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a2eb944e5d24318194a98021ea8b7ff03d0abf3d
--- /dev/null
+++ b/outputs/2025-12-13/02-25-00/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/02-25-00/.hydra/hydra.yaml b/outputs/2025-12-13/02-25-00/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ee585d81e4d88ec9eb47aba0d14bd881cfd55a2b
--- /dev/null
+++ b/outputs/2025-12-13/02-25-00/.hydra/hydra.yaml
@@ -0,0 +1,174 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+  job:
+    name: train
+    chdir: null
+    override_dirname: model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/02-25-00
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/02-25-00/.hydra/overrides.yaml b/outputs/2025-12-13/02-25-00/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f65daf34a2f9a0541f5a8001217c27417a401df6
--- /dev/null
+++ b/outputs/2025-12-13/02-25-00/.hydra/overrides.yaml
@@ -0,0 +1,5 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
diff --git a/outputs/2025-12-13/02-25-00/train.log b/outputs/2025-12-13/02-25-00/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/02-26-54/.hydra/config.yaml b/outputs/2025-12-13/02-26-54/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a2eb944e5d24318194a98021ea8b7ff03d0abf3d
--- /dev/null
+++ b/outputs/2025-12-13/02-26-54/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/02-26-54/.hydra/hydra.yaml b/outputs/2025-12-13/02-26-54/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f4b2217a4e594f4f25b8f01143cc62fa1d77c2f8
--- /dev/null
+++ b/outputs/2025-12-13/02-26-54/.hydra/hydra.yaml
@@ -0,0 +1,174 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+  job:
+    name: train
+    chdir: null
+    override_dirname: model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/02-26-54
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/02-26-54/.hydra/overrides.yaml b/outputs/2025-12-13/02-26-54/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f65daf34a2f9a0541f5a8001217c27417a401df6
--- /dev/null
+++ b/outputs/2025-12-13/02-26-54/.hydra/overrides.yaml
@@ -0,0 +1,5 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
diff --git a/outputs/2025-12-13/02-26-54/train.log b/outputs/2025-12-13/02-26-54/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/02-34-16/.hydra/config.yaml b/outputs/2025-12-13/02-34-16/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..786a71c323cbed30ab9a900256aa930e23e64e9b
--- /dev/null
+++ b/outputs/2025-12-13/02-34-16/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/02-34-16/.hydra/hydra.yaml b/outputs/2025-12-13/02-34-16/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9770e735c4e17a26666c25edfee7dcd9c5802169
--- /dev/null
+++ b/outputs/2025-12-13/02-34-16/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/02-34-16
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/02-34-16/.hydra/overrides.yaml b/outputs/2025-12-13/02-34-16/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9b0c55fce8f8c952fb4742bff85d17e07f80c09a
--- /dev/null
+++ b/outputs/2025-12-13/02-34-16/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/02-34-16/train.log b/outputs/2025-12-13/02-34-16/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/05-25-14/.hydra/config.yaml b/outputs/2025-12-13/05-25-14/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..509c142711f1c34ff281ddea9fce957a15b24e8b
--- /dev/null
+++ b/outputs/2025-12-13/05-25-14/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/05-25-14/.hydra/hydra.yaml b/outputs/2025-12-13/05-25-14/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..143c8b6ff2e26949649224fc25fb38579ded67f0
--- /dev/null
+++ b/outputs/2025-12-13/05-25-14/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/05-25-14
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/05-25-14/.hydra/overrides.yaml b/outputs/2025-12-13/05-25-14/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a41bd5ad91568893fcda7023757c68dadfa3328f
--- /dev/null
+++ b/outputs/2025-12-13/05-25-14/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/05-25-14/train.log b/outputs/2025-12-13/05-25-14/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/08-03-32/.hydra/config.yaml b/outputs/2025-12-13/08-03-32/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7c949e88740b2cc363a40e2b269c73a16565febc
--- /dev/null
+++ b/outputs/2025-12-13/08-03-32/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/08-03-32/.hydra/hydra.yaml b/outputs/2025-12-13/08-03-32/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1e0e497bc6011c727e272f6903de35d6943f1114
--- /dev/null
+++ b/outputs/2025-12-13/08-03-32/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=4
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=4,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/08-03-32
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/08-03-32/.hydra/overrides.yaml b/outputs/2025-12-13/08-03-32/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3fa08b609cc39abb3bfb4b9102ddb9fec2fddb57
--- /dev/null
+++ b/outputs/2025-12-13/08-03-32/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=4
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/08-03-32/train.log b/outputs/2025-12-13/08-03-32/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/08-05-06/.hydra/config.yaml b/outputs/2025-12-13/08-05-06/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..509c142711f1c34ff281ddea9fce957a15b24e8b
--- /dev/null
+++ b/outputs/2025-12-13/08-05-06/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/08-05-06/.hydra/hydra.yaml b/outputs/2025-12-13/08-05-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c6eb14f0824b333a056bf78ced7f1ab5c01bff69
--- /dev/null
+++ b/outputs/2025-12-13/08-05-06/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/08-05-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/08-05-06/.hydra/overrides.yaml b/outputs/2025-12-13/08-05-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a41bd5ad91568893fcda7023757c68dadfa3328f
--- /dev/null
+++ b/outputs/2025-12-13/08-05-06/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/08-05-06/train.log b/outputs/2025-12-13/08-05-06/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/08-48-55/.hydra/config.yaml b/outputs/2025-12-13/08-48-55/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..509c142711f1c34ff281ddea9fce957a15b24e8b
--- /dev/null
+++ b/outputs/2025-12-13/08-48-55/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: sudoku-4x4-frommlp-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/08-48-55/.hydra/hydra.yaml b/outputs/2025-12-13/08-48-55/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0c8df34c459954f9ee918fa956a5920ec53530ad
--- /dev/null
+++ b/outputs/2025-12-13/08-48-55/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+    - trainer.experiment_name=sudoku-4x4-frommlp-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-frommlp-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/08-48-55
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/08-48-55/.hydra/overrides.yaml b/outputs/2025-12-13/08-48-55/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a41bd5ad91568893fcda7023757c68dadfa3328f
--- /dev/null
+++ b/outputs/2025-12-13/08-48-55/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_sudoku4x4_nohint
+- trainer.experiment_name=sudoku-4x4-frommlp-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/08-48-55/train.log b/outputs/2025-12-13/08-48-55/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/11-24-41/.hydra/config.yaml b/outputs/2025-12-13/11-24-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a2c00e51332ca947eb6fd23844fbc5cd6a133152
--- /dev/null
+++ b/outputs/2025-12-13/11-24-41/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.4
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/11-24-41/.hydra/hydra.yaml b/outputs/2025-12-13/11-24-41/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..544e3c8dbfc697ac5434a0b82360c2fc59e79fa6
--- /dev/null
+++ b/outputs/2025-12-13/11-24-41/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/11-24-41
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/11-24-41/.hydra/overrides.yaml b/outputs/2025-12-13/11-24-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1ea83494aa2a9c316b02ce67a80706cafc24b814
--- /dev/null
+++ b/outputs/2025-12-13/11-24-41/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/11-24-41/train.log b/outputs/2025-12-13/11-24-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/11-38-06/.hydra/config.yaml b/outputs/2025-12-13/11-38-06/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e897975eb14218215aed56f9c638c98a29e1702e
--- /dev/null
+++ b/outputs/2025-12-13/11-38-06/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: sudoku-4x4-nohint
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 16
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/11-38-06/.hydra/hydra.yaml b/outputs/2025-12-13/11-38-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3996a0b0149db44aeed00b35c4106eb8b181d41d
--- /dev/null
+++ b/outputs/2025-12-13/11-38-06/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=sudoku-4x4-nohint
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=sudoku-4x4-nohint,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _11_sudoku
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/11-38-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/11-38-06/.hydra/overrides.yaml b/outputs/2025-12-13/11-38-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1ea83494aa2a9c316b02ce67a80706cafc24b814
--- /dev/null
+++ b/outputs/2025-12-13/11-38-06/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=sudoku-4x4-nohint
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/11-38-06/train.log b/outputs/2025-12-13/11-38-06/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/21-30-26/.hydra/config.yaml b/outputs/2025-12-13/21-30-26/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9ac091e09275e3622895d96d95386bba83af54c7
--- /dev/null
+++ b/outputs/2025-12-13/21-30-26/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048-frommlp-windows5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/21-30-26/.hydra/hydra.yaml b/outputs/2025-12-13/21-30-26/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7107298babcad82149f1d29c01a9347702543741
--- /dev/null
+++ b/outputs/2025-12-13/21-30-26/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+    - trainer.experiment_name=game_2048-frommlp-windows5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/21-30-26
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/21-30-26/.hydra/overrides.yaml b/outputs/2025-12-13/21-30-26/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0fc63e1a59e022e800bc29ca7368187127fcc701
--- /dev/null
+++ b/outputs/2025-12-13/21-30-26/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+- trainer.experiment_name=game_2048-frommlp-windows5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/21-30-26/train.log b/outputs/2025-12-13/21-30-26/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/21-31-30/.hydra/config.yaml b/outputs/2025-12-13/21-31-30/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b475536035175faaf6036d3df942cfdad833da7b
--- /dev/null
+++ b/outputs/2025-12-13/21-31-30/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/21-31-30/.hydra/hydra.yaml b/outputs/2025-12-13/21-31-30/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b66f1feba4d7130d66eed621729fa3844ef1e27f
--- /dev/null
+++ b/outputs/2025-12-13/21-31-30/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5
+    - trainer.experiment_name=game_2048-frommlp-windows5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/21-31-30
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/21-31-30/.hydra/overrides.yaml b/outputs/2025-12-13/21-31-30/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..923d6e40d675db616dc8460653baa5165a727cc0
--- /dev/null
+++ b/outputs/2025-12-13/21-31-30/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5
+- trainer.experiment_name=game_2048-frommlp-windows5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/21-31-30/train.log b/outputs/2025-12-13/21-31-30/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-13/21-32-13/.hydra/config.yaml b/outputs/2025-12-13/21-32-13/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b475536035175faaf6036d3df942cfdad833da7b
--- /dev/null
+++ b/outputs/2025-12-13/21-32-13/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8096
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 300
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 4096
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 200
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-13/21-32-13/.hydra/hydra.yaml b/outputs/2025-12-13/21-32-13/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..14154a0a622d527655c97e9231b349a375a324cc
--- /dev/null
+++ b/outputs/2025-12-13/21-32-13/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5
+    - trainer.experiment_name=game_2048-frommlp-windows5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-13/21-32-13
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-13/21-32-13/.hydra/overrides.yaml b/outputs/2025-12-13/21-32-13/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..923d6e40d675db616dc8460653baa5165a727cc0
--- /dev/null
+++ b/outputs/2025-12-13/21-32-13/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5
+- trainer.experiment_name=game_2048-frommlp-windows5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-13/21-32-13/train.log b/outputs/2025-12-13/21-32-13/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-14/09-16-21/.hydra/config.yaml b/outputs/2025-12-14/09-16-21/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d3bda31d436cc02da283185cc1291aaeb3acd225
--- /dev/null
+++ b/outputs/2025-12-14/09-16-21/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-14/09-16-21/.hydra/hydra.yaml b/outputs/2025-12-14/09-16-21/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e0edcf9e6cffa79c75865361b81f0ca402008a4c
--- /dev/null
+++ b/outputs/2025-12-14/09-16-21/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-14/09-16-21
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-14/09-16-21/.hydra/overrides.yaml b/outputs/2025-12-14/09-16-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..33387123312b7989d1ff26a08b26fc3e424d9e20
--- /dev/null
+++ b/outputs/2025-12-14/09-16-21/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-14/09-16-21/train.log b/outputs/2025-12-14/09-16-21/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-14/09-51-45/.hydra/config.yaml b/outputs/2025-12-14/09-51-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0474114594c42b1f214d6fbd557c14010e14af58
--- /dev/null
+++ b/outputs/2025-12-14/09-51-45/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-14/09-51-45/.hydra/hydra.yaml b/outputs/2025-12-14/09-51-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0811a26aade25187c7671430bd8de946d432c298
--- /dev/null
+++ b/outputs/2025-12-14/09-51-45/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-14/09-51-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-14/09-51-45/.hydra/overrides.yaml b/outputs/2025-12-14/09-51-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..714d8a78820c011a0da6fd046e623841539e386f
--- /dev/null
+++ b/outputs/2025-12-14/09-51-45/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-14/09-51-45/train.log b/outputs/2025-12-14/09-51-45/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-14/10-09-42/.hydra/config.yaml b/outputs/2025-12-14/10-09-42/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0a91346a19ceea9119541b497b055cb5fb6b6962
--- /dev/null
+++ b/outputs/2025-12-14/10-09-42/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-14/10-09-42/.hydra/hydra.yaml b/outputs/2025-12-14/10-09-42/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a150aa94bb653d0548a9578f945c6dc4265842c8
--- /dev/null
+++ b/outputs/2025-12-14/10-09-42/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-14/10-09-42
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-14/10-09-42/.hydra/overrides.yaml b/outputs/2025-12-14/10-09-42/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..38b63eb933ee771de63aa34151fdf9b483002a43
--- /dev/null
+++ b/outputs/2025-12-14/10-09-42/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-14/10-09-42/train.log b/outputs/2025-12-14/10-09-42/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-14/15-59-09/.hydra/config.yaml b/outputs/2025-12-14/15-59-09/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b1196675e2820f279f1195ae0d8b1ed3acc5f4ae
--- /dev/null
+++ b/outputs/2025-12-14/15-59-09/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-14/15-59-09/.hydra/hydra.yaml b/outputs/2025-12-14/15-59-09/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..695d60e79f1fe26aeb7bfec489b2ff2c8007ab79
--- /dev/null
+++ b/outputs/2025-12-14/15-59-09/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-14/15-59-09
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-14/15-59-09/.hydra/overrides.yaml b/outputs/2025-12-14/15-59-09/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b6f36b03291d3b93f44c6ba76557ad5926933364
--- /dev/null
+++ b/outputs/2025-12-14/15-59-09/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-14/15-59-09/train.log b/outputs/2025-12-14/15-59-09/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-14/16-44-40/.hydra/config.yaml b/outputs/2025-12-14/16-44-40/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b1196675e2820f279f1195ae0d8b1ed3acc5f4ae
--- /dev/null
+++ b/outputs/2025-12-14/16-44-40/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-14/16-44-40/.hydra/hydra.yaml b/outputs/2025-12-14/16-44-40/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9e4190c233470063a7217d314f88f61d4140ecd4
--- /dev/null
+++ b/outputs/2025-12-14/16-44-40/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-14/16-44-40
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-14/16-44-40/.hydra/overrides.yaml b/outputs/2025-12-14/16-44-40/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b6f36b03291d3b93f44c6ba76557ad5926933364
--- /dev/null
+++ b/outputs/2025-12-14/16-44-40/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-14/16-44-40/train.log b/outputs/2025-12-14/16-44-40/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-14/23-52-08/.hydra/config.yaml b/outputs/2025-12-14/23-52-08/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0a91346a19ceea9119541b497b055cb5fb6b6962
--- /dev/null
+++ b/outputs/2025-12-14/23-52-08/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-14/23-52-08/.hydra/hydra.yaml b/outputs/2025-12-14/23-52-08/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a09644e140a025cd03a418e437aa6fdd0271d6c5
--- /dev/null
+++ b/outputs/2025-12-14/23-52-08/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-14/23-52-08
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-14/23-52-08/.hydra/overrides.yaml b/outputs/2025-12-14/23-52-08/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..38b63eb933ee771de63aa34151fdf9b483002a43
--- /dev/null
+++ b/outputs/2025-12-14/23-52-08/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-14/23-52-08/train.log b/outputs/2025-12-14/23-52-08/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/09-04-57/.hydra/config.yaml b/outputs/2025-12-16/09-04-57/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6a2c60df191233770912f9a765a5e228feb6c75b
--- /dev/null
+++ b/outputs/2025-12-16/09-04-57/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score-3epoch
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/09-04-57/.hydra/hydra.yaml b/outputs/2025-12-16/09-04-57/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..eba9cae8aaf210a649d0d384c0072ec71a87cf93
--- /dev/null
+++ b/outputs/2025-12-16/09-04-57/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score-3epoch
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score-3epoch,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/09-04-57
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/09-04-57/.hydra/overrides.yaml b/outputs/2025-12-16/09-04-57/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..616aacadbb964bb858b6b61765f421cf6605052f
--- /dev/null
+++ b/outputs/2025-12-16/09-04-57/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score-3epoch
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/09-04-57/train.log b/outputs/2025-12-16/09-04-57/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/09-59-25/.hydra/config.yaml b/outputs/2025-12-16/09-59-25/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b1196675e2820f279f1195ae0d8b1ed3acc5f4ae
--- /dev/null
+++ b/outputs/2025-12-16/09-59-25/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/09-59-25/.hydra/hydra.yaml b/outputs/2025-12-16/09-59-25/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d4dff70b279ee163caba7cc17f111725190c09e6
--- /dev/null
+++ b/outputs/2025-12-16/09-59-25/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/09-59-25
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/09-59-25/.hydra/overrides.yaml b/outputs/2025-12-16/09-59-25/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b6f36b03291d3b93f44c6ba76557ad5926933364
--- /dev/null
+++ b/outputs/2025-12-16/09-59-25/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/09-59-25/train.log b/outputs/2025-12-16/09-59-25/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/14-22-43/.hydra/config.yaml b/outputs/2025-12-16/14-22-43/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0a91346a19ceea9119541b497b055cb5fb6b6962
--- /dev/null
+++ b/outputs/2025-12-16/14-22-43/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/14-22-43/.hydra/hydra.yaml b/outputs/2025-12-16/14-22-43/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9c4968fe7a4e581f62e3e6bd421523a53e48486e
--- /dev/null
+++ b/outputs/2025-12-16/14-22-43/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/14-22-43
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/14-22-43/.hydra/overrides.yaml b/outputs/2025-12-16/14-22-43/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..38b63eb933ee771de63aa34151fdf9b483002a43
--- /dev/null
+++ b/outputs/2025-12-16/14-22-43/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/14-22-43/train.log b/outputs/2025-12-16/14-22-43/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/17-03-37/.hydra/config.yaml b/outputs/2025-12-16/17-03-37/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..590d6d41f9cd28f33946ecfde52027d02cb6f261
--- /dev/null
+++ b/outputs/2025-12-16/17-03-37/.hydra/config.yaml
@@ -0,0 +1,864 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 14400
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-box2-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: 'You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets You are provided with a symbol grid and the
+      zero-indexed coordinates of the player, each box, and each target. Coordinates
+      range from the top-left corner (0, 0) to the bottom-right corner (5, 5). When
+      you are exactly next to a box, you can push it by moving in the same direction.
+      You cannot push a box through a wall, and you cannot pull a box. The answer
+      should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+      A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/17-03-37/.hydra/hydra.yaml b/outputs/2025-12-16/17-03-37/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..983522fba5db8f418fa299c88194b90af53dd49b
--- /dev/null
+++ b/outputs/2025-12-16/17-03-37/.hydra/hydra.yaml
@@ -0,0 +1,204 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+    - custom_envs.SimpleSokoban.env_config.num_boxes=2
+    - 'custom_envs.SimpleSokoban.env_instruction=''You are solving the Sokoban puzzle.
+      You are the player and you need to push all boxes to targets You are provided
+      with a symbol grid and the zero-indexed coordinates of the player, each box,
+      and each target. Coordinates range from the top-left corner (0, 0) to the bottom-right
+      corner (5, 5). When you are exactly next to a box, you can push it by moving
+      in the same direction. You cannot push a box through a wall, and you cannot
+      pull a box. The answer should be a sequence of actions, like <answer>Right ||
+      Right || Up</answer>. A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'''
+    - trainer.experiment_name=sokoban-box2-starpos-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.SimpleSokoban.env_config.num_boxes=2,custom_envs.SimpleSokoban.env_instruction=''You
+      are solving the Sokoban puzzle. You are the player and you need to push all
+      boxes to targets You are provided with a symbol grid and the zero-indexed coordinates
+      of the player, each box, and each target. Coordinates range from the top-left
+      corner (0, 0) to the bottom-right corner (5, 5). When you are exactly next to
+      a box, you can push it by moving in the same direction. You cannot push a box
+      through a wall, and you cannot pull a box. The answer should be a sequence of
+      actions, like <answer>Right || Right || Up</answer>. A sample full output is
+      as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'',model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470,system.CUDA_VISIBLE_DEVICES=''0'',trainer.experiment_name=sokoban-box2-starpos-SPA,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/17-03-37
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/17-03-37/.hydra/overrides.yaml b/outputs/2025-12-16/17-03-37/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d781124335e0b7fffb16d4cad25969bedea9d646
--- /dev/null
+++ b/outputs/2025-12-16/17-03-37/.hydra/overrides.yaml
@@ -0,0 +1,23 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+- custom_envs.SimpleSokoban.env_config.num_boxes=2
+- 'custom_envs.SimpleSokoban.env_instruction=''You are solving the Sokoban puzzle.
+  You are the player and you need to push all boxes to targets You are provided with
+  a symbol grid and the zero-indexed coordinates of the player, each box, and each
+  target. Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+  (5, 5). When you are exactly next to a box, you can push it by moving in the same
+  direction. You cannot push a box through a wall, and you cannot pull a box. The
+  answer should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+  A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+  (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation> 1 Down
+  – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left of box, ready
+  to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+  </think> <answer> Down || Left || Down </answer>'''
+- trainer.experiment_name=sokoban-box2-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/17-06-16/.hydra/config.yaml b/outputs/2025-12-16/17-06-16/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..04f0b19075499161b83426aeb54ecc38b0c7f606
--- /dev/null
+++ b/outputs/2025-12-16/17-06-16/.hydra/config.yaml
@@ -0,0 +1,864 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-box2-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: 'You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets You are provided with a symbol grid and the
+      zero-indexed coordinates of the player, each box, and each target. Coordinates
+      range from the top-left corner (0, 0) to the bottom-right corner (5, 5). When
+      you are exactly next to a box, you can push it by moving in the same direction.
+      You cannot push a box through a wall, and you cannot pull a box. The answer
+      should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+      A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/17-06-16/.hydra/hydra.yaml b/outputs/2025-12-16/17-06-16/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6e63b61df9204798069a649ab71e79e71f9436ad
--- /dev/null
+++ b/outputs/2025-12-16/17-06-16/.hydra/hydra.yaml
@@ -0,0 +1,204 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+    - custom_envs.SimpleSokoban.env_config.num_boxes=2
+    - 'custom_envs.SimpleSokoban.env_instruction=''You are solving the Sokoban puzzle.
+      You are the player and you need to push all boxes to targets You are provided
+      with a symbol grid and the zero-indexed coordinates of the player, each box,
+      and each target. Coordinates range from the top-left corner (0, 0) to the bottom-right
+      corner (5, 5). When you are exactly next to a box, you can push it by moving
+      in the same direction. You cannot push a box through a wall, and you cannot
+      pull a box. The answer should be a sequence of actions, like <answer>Right ||
+      Right || Up</answer>. A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'''
+    - trainer.experiment_name=sokoban-box2-starpos-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.SimpleSokoban.env_config.num_boxes=2,custom_envs.SimpleSokoban.env_instruction=''You
+      are solving the Sokoban puzzle. You are the player and you need to push all
+      boxes to targets You are provided with a symbol grid and the zero-indexed coordinates
+      of the player, each box, and each target. Coordinates range from the top-left
+      corner (0, 0) to the bottom-right corner (5, 5). When you are exactly next to
+      a box, you can push it by moving in the same direction. You cannot push a box
+      through a wall, and you cannot pull a box. The answer should be a sequence of
+      actions, like <answer>Right || Right || Up</answer>. A sample full output is
+      as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'',model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470,system.CUDA_VISIBLE_DEVICES=''0'',trainer.experiment_name=sokoban-box2-starpos-SPA,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/17-06-16
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/17-06-16/.hydra/overrides.yaml b/outputs/2025-12-16/17-06-16/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d781124335e0b7fffb16d4cad25969bedea9d646
--- /dev/null
+++ b/outputs/2025-12-16/17-06-16/.hydra/overrides.yaml
@@ -0,0 +1,23 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+- custom_envs.SimpleSokoban.env_config.num_boxes=2
+- 'custom_envs.SimpleSokoban.env_instruction=''You are solving the Sokoban puzzle.
+  You are the player and you need to push all boxes to targets You are provided with
+  a symbol grid and the zero-indexed coordinates of the player, each box, and each
+  target. Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+  (5, 5). When you are exactly next to a box, you can push it by moving in the same
+  direction. You cannot push a box through a wall, and you cannot pull a box. The
+  answer should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+  A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+  (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation> 1 Down
+  – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left of box, ready
+  to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+  </think> <answer> Down || Left || Down </answer>'''
+- trainer.experiment_name=sokoban-box2-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/17-06-16/train.log b/outputs/2025-12-16/17-06-16/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/17-11-13/.hydra/config.yaml b/outputs/2025-12-16/17-11-13/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..de13268c99d512fc672787d8355fe45c1b91e387
--- /dev/null
+++ b/outputs/2025-12-16/17-11-13/.hydra/config.yaml
@@ -0,0 +1,861 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-box2-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: 'You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets You are provided with a symbol grid and the
+      zero-indexed coordinates of the player, each box, and each target. Coordinates
+      range from the top-left corner (0, 0) to the bottom-right corner (5, 5). When
+      you are exactly next to a box, you can push it by moving in the same direction.
+      You cannot push a box through a wall, and you cannot pull a box. The answer
+      should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+      A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/17-11-13/.hydra/hydra.yaml b/outputs/2025-12-16/17-11-13/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d02c40c36be872302827f972a09d65547fc30ee1
--- /dev/null
+++ b/outputs/2025-12-16/17-11-13/.hydra/hydra.yaml
@@ -0,0 +1,204 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+    - custom_envs.CoordSokoban.env_config.num_boxes=2
+    - 'custom_envs.CoordSokoban.env_instruction=''You are solving the Sokoban puzzle.
+      You are the player and you need to push all boxes to targets You are provided
+      with a symbol grid and the zero-indexed coordinates of the player, each box,
+      and each target. Coordinates range from the top-left corner (0, 0) to the bottom-right
+      corner (5, 5). When you are exactly next to a box, you can push it by moving
+      in the same direction. You cannot push a box through a wall, and you cannot
+      pull a box. The answer should be a sequence of actions, like <answer>Right ||
+      Right || Up</answer>. A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'''
+    - trainer.experiment_name=sokoban-box2-starpos-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.CoordSokoban.env_config.num_boxes=2,custom_envs.CoordSokoban.env_instruction=''You
+      are solving the Sokoban puzzle. You are the player and you need to push all
+      boxes to targets You are provided with a symbol grid and the zero-indexed coordinates
+      of the player, each box, and each target. Coordinates range from the top-left
+      corner (0, 0) to the bottom-right corner (5, 5). When you are exactly next to
+      a box, you can push it by moving in the same direction. You cannot push a box
+      through a wall, and you cannot pull a box. The answer should be a sequence of
+      actions, like <answer>Right || Right || Up</answer>. A sample full output is
+      as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'',model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470,system.CUDA_VISIBLE_DEVICES=''0'',trainer.experiment_name=sokoban-box2-starpos-SPA,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/17-11-13
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/17-11-13/.hydra/overrides.yaml b/outputs/2025-12-16/17-11-13/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b180acfe3df59acfa8b06ff555af2bab4c57df85
--- /dev/null
+++ b/outputs/2025-12-16/17-11-13/.hydra/overrides.yaml
@@ -0,0 +1,23 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+- custom_envs.CoordSokoban.env_config.num_boxes=2
+- 'custom_envs.CoordSokoban.env_instruction=''You are solving the Sokoban puzzle.
+  You are the player and you need to push all boxes to targets You are provided with
+  a symbol grid and the zero-indexed coordinates of the player, each box, and each
+  target. Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+  (5, 5). When you are exactly next to a box, you can push it by moving in the same
+  direction. You cannot push a box through a wall, and you cannot pull a box. The
+  answer should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+  A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+  (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation> 1 Down
+  – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left of box, ready
+  to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+  </think> <answer> Down || Left || Down </answer>'''
+- trainer.experiment_name=sokoban-box2-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/17-11-13/train.log b/outputs/2025-12-16/17-11-13/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/17-14-54/.hydra/config.yaml b/outputs/2025-12-16/17-14-54/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e5a0fcccd4f186790d4d2ae38c3fe78c5a2e3fa7
--- /dev/null
+++ b/outputs/2025-12-16/17-14-54/.hydra/config.yaml
@@ -0,0 +1,861 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-box2-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: 'You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets You are provided with a symbol grid and the
+      zero-indexed coordinates of the player, each box, and each target. Coordinates
+      range from the top-left corner (0, 0) to the bottom-right corner (5, 5). When
+      you are exactly next to a box, you can push it by moving in the same direction.
+      You cannot push a box through a wall, and you cannot pull a box. The answer
+      should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+      A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/17-14-54/.hydra/hydra.yaml b/outputs/2025-12-16/17-14-54/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bd607427f0b2a360cbf09ff82ba3c3aff3f183f9
--- /dev/null
+++ b/outputs/2025-12-16/17-14-54/.hydra/hydra.yaml
@@ -0,0 +1,204 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+    - custom_envs.CoordSokoban.env_config.num_boxes=2
+    - 'custom_envs.CoordSokoban.env_instruction=''You are solving the Sokoban puzzle.
+      You are the player and you need to push all boxes to targets You are provided
+      with a symbol grid and the zero-indexed coordinates of the player, each box,
+      and each target. Coordinates range from the top-left corner (0, 0) to the bottom-right
+      corner (5, 5). When you are exactly next to a box, you can push it by moving
+      in the same direction. You cannot push a box through a wall, and you cannot
+      pull a box. The answer should be a sequence of actions, like <answer>Right ||
+      Right || Up</answer>. A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'''
+    - trainer.experiment_name=sokoban-box2-starpos-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.CoordSokoban.env_config.num_boxes=2,custom_envs.CoordSokoban.env_instruction=''You
+      are solving the Sokoban puzzle. You are the player and you need to push all
+      boxes to targets You are provided with a symbol grid and the zero-indexed coordinates
+      of the player, each box, and each target. Coordinates range from the top-left
+      corner (0, 0) to the bottom-right corner (5, 5). When you are exactly next to
+      a box, you can push it by moving in the same direction. You cannot push a box
+      through a wall, and you cannot pull a box. The answer should be a sequence of
+      actions, like <answer>Right || Right || Up</answer>. A sample full output is
+      as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'',model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=sokoban-box2-starpos-SPA,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _2_sokoban
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/17-14-54
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/17-14-54/.hydra/overrides.yaml b/outputs/2025-12-16/17-14-54/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..68d5c59ef52d99373d33ea7c9fc76b957ac9c122
--- /dev/null
+++ b/outputs/2025-12-16/17-14-54/.hydra/overrides.yaml
@@ -0,0 +1,23 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+- custom_envs.CoordSokoban.env_config.num_boxes=2
+- 'custom_envs.CoordSokoban.env_instruction=''You are solving the Sokoban puzzle.
+  You are the player and you need to push all boxes to targets You are provided with
+  a symbol grid and the zero-indexed coordinates of the player, each box, and each
+  target. Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+  (5, 5). When you are exactly next to a box, you can push it by moving in the same
+  direction. You cannot push a box through a wall, and you cannot pull a box. The
+  answer should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+  A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+  (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation> 1 Down
+  – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left of box, ready
+  to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+  </think> <answer> Down || Left || Down </answer>'''
+- trainer.experiment_name=sokoban-box2-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/17-14-54/train.log b/outputs/2025-12-16/17-14-54/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/18-56-32/.hydra/config.yaml b/outputs/2025-12-16/18-56-32/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8775ecebc7f5ef8749e9d9fb95736c7bb7390f80
--- /dev/null
+++ b/outputs/2025-12-16/18-56-32/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/18-56-32/.hydra/overrides.yaml b/outputs/2025-12-16/18-56-32/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fdf7373fe4af122d83bef5424848d49d5a5eaa2d
--- /dev/null
+++ b/outputs/2025-12-16/18-56-32/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/18-56-32/train.log b/outputs/2025-12-16/18-56-32/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/18-57-51/.hydra/config.yaml b/outputs/2025-12-16/18-57-51/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8775ecebc7f5ef8749e9d9fb95736c7bb7390f80
--- /dev/null
+++ b/outputs/2025-12-16/18-57-51/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/18-57-51/.hydra/hydra.yaml b/outputs/2025-12-16/18-57-51/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..13652b8f46e9aa6c12dee4fd8978ca80a19568a5
--- /dev/null
+++ b/outputs/2025-12-16/18-57-51/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/18-57-51
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/18-57-51/.hydra/overrides.yaml b/outputs/2025-12-16/18-57-51/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fdf7373fe4af122d83bef5424848d49d5a5eaa2d
--- /dev/null
+++ b/outputs/2025-12-16/18-57-51/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/18-57-51/train.log b/outputs/2025-12-16/18-57-51/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/19-51-25/.hydra/config.yaml b/outputs/2025-12-16/19-51-25/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e381537b13554e426308bbf00a8401005678b846
--- /dev/null
+++ b/outputs/2025-12-16/19-51-25/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '6'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/19-51-25/.hydra/hydra.yaml b/outputs/2025-12-16/19-51-25/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9b3f568b1d044084f963fed9ac7a39a7c16734c1
--- /dev/null
+++ b/outputs/2025-12-16/19-51-25/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='6'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648,system.CUDA_VISIBLE_DEVICES='6',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/19-51-25
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/19-51-25/.hydra/overrides.yaml b/outputs/2025-12-16/19-51-25/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d90fda331eb51741abbb515848c295a7b61052a7
--- /dev/null
+++ b/outputs/2025-12-16/19-51-25/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/19-51-25/train.log b/outputs/2025-12-16/19-51-25/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/19-54-46/.hydra/hydra.yaml b/outputs/2025-12-16/19-54-46/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..43ef8da6ce053e0383f26ba23540d5752b346ead
--- /dev/null
+++ b/outputs/2025-12-16/19-54-46/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='6'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648,system.CUDA_VISIBLE_DEVICES='6',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/19-54-46
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/19-54-46/.hydra/overrides.yaml b/outputs/2025-12-16/19-54-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d90fda331eb51741abbb515848c295a7b61052a7
--- /dev/null
+++ b/outputs/2025-12-16/19-54-46/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/19-54-46/train.log b/outputs/2025-12-16/19-54-46/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/19-58-16/.hydra/config.yaml b/outputs/2025-12-16/19-58-16/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e381537b13554e426308bbf00a8401005678b846
--- /dev/null
+++ b/outputs/2025-12-16/19-58-16/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '6'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/19-58-16/.hydra/hydra.yaml b/outputs/2025-12-16/19-58-16/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5a15dd570025a56f7eb11564ef079dd2023ddbe5
--- /dev/null
+++ b/outputs/2025-12-16/19-58-16/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='6'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648,system.CUDA_VISIBLE_DEVICES='6',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/19-58-16
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/19-58-16/.hydra/overrides.yaml b/outputs/2025-12-16/19-58-16/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d90fda331eb51741abbb515848c295a7b61052a7
--- /dev/null
+++ b/outputs/2025-12-16/19-58-16/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/20-08-43/.hydra/config.yaml b/outputs/2025-12-16/20-08-43/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e381537b13554e426308bbf00a8401005678b846
--- /dev/null
+++ b/outputs/2025-12-16/20-08-43/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '6'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/20-08-43/.hydra/hydra.yaml b/outputs/2025-12-16/20-08-43/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e446585fefad4990337b80a40624cd240fff2975
--- /dev/null
+++ b/outputs/2025-12-16/20-08-43/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='6'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648,system.CUDA_VISIBLE_DEVICES='6',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/20-08-43
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/20-08-43/.hydra/overrides.yaml b/outputs/2025-12-16/20-08-43/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d90fda331eb51741abbb515848c295a7b61052a7
--- /dev/null
+++ b/outputs/2025-12-16/20-08-43/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/20-08-43/train.log b/outputs/2025-12-16/20-08-43/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/20-10-32/.hydra/overrides.yaml b/outputs/2025-12-16/20-10-32/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d90fda331eb51741abbb515848c295a7b61052a7
--- /dev/null
+++ b/outputs/2025-12-16/20-10-32/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='6'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+- trainer.experiment_name=game_2048-frommlp-windows5-7000score
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-16/20-10-32/train.log b/outputs/2025-12-16/20-10-32/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/20-22-19/.hydra/hydra.yaml b/outputs/2025-12-16/20-22-19/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b632045a1317bdfcaf12c47e823c9232f3cca7ca
--- /dev/null
+++ b/outputs/2025-12-16/20-22-19/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+    - trainer.experiment_name=game_2048-frommlp-windows5-7000score
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5-7000score,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-16/20-22-19
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-16/20-22-19/train.log b/outputs/2025-12-16/20-22-19/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-16/20-26-33/.hydra/config.yaml b/outputs/2025-12-16/20-26-33/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e381537b13554e426308bbf00a8401005678b846
--- /dev/null
+++ b/outputs/2025-12-16/20-26-33/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 8192
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: game_2048-frommlp-windows5-7000score
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '6'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_2048_windows5_7000scores/checkpoint-3648
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-16/20-48-11/train.log b/outputs/2025-12-16/20-48-11/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-17/01-38-45/.hydra/config.yaml b/outputs/2025-12-17/01-38-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4ebe5ddeea336a75726392ec14c80b904c1b4b8d
--- /dev/null
+++ b/outputs/2025-12-17/01-38-45/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: frozenlake-static-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_3_frozen_lake-0.5B-text_with_coordinates-qwen/global_step_3068
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-17/01-38-45/train.log b/outputs/2025-12-17/01-38-45/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-17/02-15-58/.hydra/config.yaml b/outputs/2025-12-17/02-15-58/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ccd232fe1bfd58cb468e81ec5de742b3a8354f18
--- /dev/null
+++ b/outputs/2025-12-17/02-15-58/.hydra/config.yaml
@@ -0,0 +1,861 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-box2-starpos-stateestimate
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: 'You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets You are provided with a symbol grid and the
+      zero-indexed coordinates of the player, each box, and each target. Coordinates
+      range from the top-left corner (0, 0) to the bottom-right corner (5, 5). When
+      you are exactly next to a box, you can push it by moving in the same direction.
+      You cannot push a box through a wall, and you cannot pull a box. The answer
+      should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+      A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-17/07-07-14/.hydra/config.yaml b/outputs/2025-12-17/07-07-14/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e5a0fcccd4f186790d4d2ae38c3fe78c5a2e3fa7
--- /dev/null
+++ b/outputs/2025-12-17/07-07-14/.hydra/config.yaml
@@ -0,0 +1,861 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: sokoban-box2-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: 'You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets You are provided with a symbol grid and the
+      zero-indexed coordinates of the player, each box, and each target. Coordinates
+      range from the top-left corner (0, 0) to the bottom-right corner (5, 5). When
+      you are exactly next to a box, you can push it by moving in the same direction.
+      You cannot push a box through a wall, and you cannot pull a box. The answer
+      should be a sequence of actions, like <answer>Right || Right || Up</answer>.
+      A sample full output is as follows: <think> <observation>######\n#_####\n#_P###\n#_X#_#\n#__O_#\n######\nPlayer
+      (P) is at (2,2); box (X) is at (3,2); target (O) is at (4,3).</observation>
+      1 Down – I push box to (4,2). 2 Left – I step to (3,1). 3 Down – I stand left
+      of box, ready to push it Right onto target. <prediction>######\n#_####\n#__###\n#__#_#\n#PXO_#\n######</prediction>
+      </think> <answer> Down || Left || Down </answer>'
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_2_sokoban-1.5B-text_with_coordinates-qwen/global_step_1470
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-17/07-07-14/train.log b/outputs/2025-12-17/07-07-14/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/06-51-46/.hydra/config.yaml b/outputs/2025-12-18/06-51-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..13be45e30aca23b3a82909a265e12259b822c885
--- /dev/null
+++ b/outputs/2025-12-18/06-51-46/.hydra/config.yaml
@@ -0,0 +1,853 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 3600
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: frozenlake-static-starpos-stateestimate
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: You are solving the FrozenLake puzzle.\nForbid the whole and
+      go to the target.\nYou may move to the unintended direction due to the slippery
+      ice.\nExample answer format:\n<think>To forbid the hole and go to the target,
+      I should go left then go up.</think>\n<answer>Left || Up</answer>\nA sample
+      full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+      at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+      2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now
+      on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+      || Left || Left</answer>
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-18/12-05-46/.hydra/config.yaml b/outputs/2025-12-18/12-05-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d53113a121e8390e3bad74691af838047071a5bb
--- /dev/null
+++ b/outputs/2025-12-18/12-05-46/.hydra/config.yaml
@@ -0,0 +1,914 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: "You are solving the FrozenLake puzzle.\nForbid the whole and\
+      \ go to the target.\nYou may move to the unintended direction due to the slippery\
+      \ ice.\nExample answer format:\n<think>To forbid the hole and go to the target,\
+      \ I should go left then go up.</think>\n<answer>Left || Up</answer>\nA sample\
+      \ full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n\
+      </observation>\nPlayer at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1\
+      \ Up move to safe ice\n(2,2). 2 Left slide to (2,1), adjacent to goal. 3 Left\
+      \ reach goal (2,0);\nplayer now on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n\
+      </prediction>\n</think> \n<answer>Up || Left || Left</answer>"
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: glm-4.6
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
diff --git a/outputs/2025-12-18/12-05-46/.hydra/overrides.yaml b/outputs/2025-12-18/12-05-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/12-05-46/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/12-22-15/.hydra/hydra.yaml b/outputs/2025-12-18/12-22-15/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9935447891cc04f3bfa964bc3f8a8e0c932d9519
--- /dev/null
+++ b/outputs/2025-12-18/12-22-15/.hydra/hydra.yaml
@@ -0,0 +1,157 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: agent_proxy
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: eval.yaml
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/12-22-15
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/13-45-32/.hydra/hydra.yaml b/outputs/2025-12-18/13-45-32/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c93d78220e762fde182b219199a6de79f406737d
--- /dev/null
+++ b/outputs/2025-12-18/13-45-32/.hydra/hydra.yaml
@@ -0,0 +1,196 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_SPA_frozenlake
+    - custom_envs.CoordFrozenLake.env_instruction='You are solving the FrozenLake
+      puzzle.\nForbid the whole and go to the target.\nYou may move to the unintended
+      direction due to the slippery ice.\nExample answer format:\n<think>To forbid
+      the hole and go to the target, I should go left then go up.</think>\n<answer>Left
+      || Up</answer>\nA sample full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+      at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+      2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now
+      on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+      || Left || Left</answer>'
+    - trainer.experiment_name=frozenlake-static-starpos-SPA-7200
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.CoordFrozenLake.env_instruction='You
+      are solving the FrozenLake puzzle.\nForbid the whole and go to the target.\nYou
+      may move to the unintended direction due to the slippery ice.\nExample answer
+      format:\n<think>To forbid the hole and go to the target, I should go left then
+      go up.</think>\n<answer>Left || Up</answer>\nA sample full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+      at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+      2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now
+      on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+      || Left || Left</answer>',model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_SPA_frozenlake,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=frozenlake-static-starpos-SPA-7200,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _3_frozen_lake
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/13-45-32
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/13-45-32/.hydra/overrides.yaml b/outputs/2025-12-18/13-45-32/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d81ef80191b9274766e0fa61ef9e91d637198deb
--- /dev/null
+++ b/outputs/2025-12-18/13-45-32/.hydra/overrides.yaml
@@ -0,0 +1,19 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_SPA_frozenlake
+- custom_envs.CoordFrozenLake.env_instruction='You are solving the FrozenLake puzzle.\nForbid
+  the whole and go to the target.\nYou may move to the unintended direction due to
+  the slippery ice.\nExample answer format:\n<think>To forbid the hole and go to the
+  target, I should go left then go up.</think>\n<answer>Left || Up</answer>\nA sample
+  full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+  at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+  2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now on
+  G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+  || Left || Left</answer>'
+- trainer.experiment_name=frozenlake-static-starpos-SPA-7200
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-18/13-45-32/train.log b/outputs/2025-12-18/13-45-32/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/14-35-03/.hydra/config.yaml b/outputs/2025-12-18/14-35-03/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d53ff4cc611b55b99dc6f9b8622230c7781a269f
--- /dev/null
+++ b/outputs/2025-12-18/14-35-03/.hydra/config.yaml
@@ -0,0 +1,853 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 1000
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: frozenlake-static-starpos-SPA-7200
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: You are solving the FrozenLake puzzle.\nForbid the whole and
+      go to the target.\nYou may move to the unintended direction due to the slippery
+      ice.\nExample answer format:\n<think>To forbid the hole and go to the target,
+      I should go left then go up.</think>\n<answer>Left || Up</answer>\nA sample
+      full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+      at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+      2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now
+      on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+      || Left || Left</answer>
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_SPA_frozenlake
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-18/14-35-03/train.log b/outputs/2025-12-18/14-35-03/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/14-41-20/.hydra/config.yaml b/outputs/2025-12-18/14-41-20/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..38b327012fd087fd7f73356c43aa33908be98b28
--- /dev/null
+++ b/outputs/2025-12-18/14-41-20/.hydra/config.yaml
@@ -0,0 +1,915 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
diff --git a/outputs/2025-12-18/14-41-20/eval_api.log b/outputs/2025-12-18/14-41-20/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..93f19b9f6a24f66766349dbe6027295dfc97267c
--- /dev/null
+++ b/outputs/2025-12-18/14-41-20/eval_api.log
@@ -0,0 +1,40 @@
+[2025-12-18 14:41:30,689][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f43293a1b70>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 35aeeb00-2ee0-453d-ae52-0844784b9f8d)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 14:41:30,690][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 14:41:42,053][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /api/resolve-cache/models/Qwen/Qwen2.5-0.5B-Instruct/7ae557604adf67be50417f59c2c2f167def9a775/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f432937f610>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 4d55ffa9-1668-4b3a-bfc8-bbba0eb34bb1)')' thrown while requesting HEAD https://huggingface.co/api/resolve-cache/models/Qwen/Qwen2.5-0.5B-Instruct/7ae557604adf67be50417f59c2c2f167def9a775/tokenizer_config.json
+[2025-12-18 14:41:42,055][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 14:41:47,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,473][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,505][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:47,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:48,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:48,771][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:48,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:48,909][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:48,925][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:49,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:49,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:49,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:49,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:49,079][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:49,088][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:49,090][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:50,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:50,140][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465848 seconds
+[2025-12-18 14:41:50,141][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393811 seconds
+[2025-12-18 14:41:50,142][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448925 seconds
+[2025-12-18 14:41:50,142][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477424 seconds
+[2025-12-18 14:41:50,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:50,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:50,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:50,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:51,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:51,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:52,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:41:52,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/14-43-23/.hydra/config.yaml b/outputs/2025-12-18/14-43-23/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..47f297e69063ff7d23586b864ba13941cd54e4d5
--- /dev/null
+++ b/outputs/2025-12-18/14-43-23/.hydra/config.yaml
@@ -0,0 +1,915 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
diff --git a/outputs/2025-12-18/14-43-23/.hydra/overrides.yaml b/outputs/2025-12-18/14-43-23/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/14-43-23/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/14-43-23/eval_api.log b/outputs/2025-12-18/14-43-23/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..d711a2bc866d3c1593d1e252e8492c244d1f7a5f
--- /dev/null
+++ b/outputs/2025-12-18/14-43-23/eval_api.log
@@ -0,0 +1,99 @@
+[2025-12-18 14:43:25,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448074 seconds
+[2025-12-18 14:43:27,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:27,967][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:28,130][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:28,134][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:28,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:28,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:28,843][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:28,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:29,048][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:30,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:30,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:30,426][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401857 seconds
+[2025-12-18 14:43:30,428][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407552 seconds
+[2025-12-18 14:43:30,429][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430952 seconds
+[2025-12-18 14:43:30,429][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478202 seconds
+[2025-12-18 14:43:30,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:30,602][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:30,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:31,189][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:31,431][openai._base_client][INFO] - Retrying request to /chat/completions in 0.786184 seconds
+[2025-12-18 14:43:31,485][openai._base_client][INFO] - Retrying request to /chat/completions in 0.868826 seconds
+[2025-12-18 14:43:31,824][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:32,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:32,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:33,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:33,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:33,536][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:33,742][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:33,905][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:33,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:34,267][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:34,372][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:34,824][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:34,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:35,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.799845 seconds
+[2025-12-18 14:43:35,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:36,072][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:36,341][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:38,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:43,422][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396944 seconds
+[2025-12-18 14:43:48,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490952 seconds
+[2025-12-18 14:43:48,825][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967289 seconds
+[2025-12-18 14:43:49,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:50,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:50,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:50,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:53,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:53,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:43:58,683][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448381 seconds
+[2025-12-18 14:43:59,692][openai._base_client][INFO] - Retrying request to /chat/completions in 0.829064 seconds
+[2025-12-18 14:44:02,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:44:10,537][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411047 seconds
+[2025-12-18 14:44:14,561][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:44:15,540][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427632 seconds
+[2025-12-18 14:44:20,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:44:25,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454970 seconds
+[2025-12-18 14:44:31,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:44:36,345][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496244 seconds
+[2025-12-18 14:44:39,752][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:44:44,762][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464400 seconds
+[2025-12-18 14:44:49,132][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:44:54,142][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430631 seconds
+[2025-12-18 14:44:59,581][openai._base_client][INFO] - Retrying request to /chat/completions in 0.827363 seconds
+[2025-12-18 14:45:10,988][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381337 seconds
+[2025-12-18 14:45:16,377][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920119 seconds
+[2025-12-18 14:45:32,319][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377376 seconds
+[2025-12-18 14:45:37,705][openai._base_client][INFO] - Retrying request to /chat/completions in 0.810609 seconds
+[2025-12-18 14:45:41,758][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:45:46,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458481 seconds
+[2025-12-18 14:45:49,932][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:45:54,941][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423418 seconds
+[2025-12-18 14:45:58,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:46:03,257][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498860 seconds
+[2025-12-18 14:46:11,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:46:16,382][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480664 seconds
+[2025-12-18 14:46:21,372][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:46:26,381][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488157 seconds
+[2025-12-18 14:46:34,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:46:39,540][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376321 seconds
+[2025-12-18 14:46:44,328][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:46:49,339][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386550 seconds
+[2025-12-18 14:46:54,636][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:46:59,645][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448864 seconds
+[2025-12-18 14:47:00,591][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944386 seconds
+[2025-12-18 14:47:04,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:47:09,935][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397175 seconds
+[2025-12-18 14:47:13,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:47:18,655][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491308 seconds
+[2025-12-18 14:47:23,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:47:28,625][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438693 seconds
+[2025-12-18 14:47:32,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:47:37,019][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411984 seconds
+[2025-12-18 14:47:40,659][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:47:45,668][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426719 seconds
+[2025-12-18 14:47:51,103][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921815 seconds
+[2025-12-18 14:47:56,270][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:48:01,279][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404212 seconds
+[2025-12-18 14:48:06,691][openai._base_client][INFO] - Retrying request to /chat/completions in 0.940018 seconds
diff --git a/outputs/2025-12-18/14-49-17/.hydra/config.yaml b/outputs/2025-12-18/14-49-17/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d28de1e7fab758543e07784afa776dc2e07f1a2c
--- /dev/null
+++ b/outputs/2025-12-18/14-49-17/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/14-49-17/.hydra/hydra.yaml b/outputs/2025-12-18/14-49-17/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..207ed90bd603a3111332ec7550c2f5b24f2152ee
--- /dev/null
+++ b/outputs/2025-12-18/14-49-17/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/14-49-17
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/14-49-17/.hydra/overrides.yaml b/outputs/2025-12-18/14-49-17/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/14-49-17/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/14-49-17/eval_api.log b/outputs/2025-12-18/14-49-17/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..cce71867b5cd209eae4c5b84b262a4f024de1531
--- /dev/null
+++ b/outputs/2025-12-18/14-49-17/eval_api.log
@@ -0,0 +1,283 @@
+[2025-12-18 14:49:22,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:22,691][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:23,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:23,190][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:23,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:24,055][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:24,217][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:24,279][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420964 seconds
+[2025-12-18 14:49:24,279][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429591 seconds
+[2025-12-18 14:49:24,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410620 seconds
+[2025-12-18 14:49:24,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497642 seconds
+[2025-12-18 14:49:24,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404138 seconds
+[2025-12-18 14:49:24,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462992 seconds
+[2025-12-18 14:49:24,282][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492146 seconds
+[2025-12-18 14:49:24,284][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400007 seconds
+[2025-12-18 14:49:25,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:25,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:25,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:26,664][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:26,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:27,461][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:27,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:27,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:27,780][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:28,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:28,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:28,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:29,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:29,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:29,325][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:29,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:29,688][openai._base_client][INFO] - Retrying request to /chat/completions in 0.768790 seconds
+[2025-12-18 14:49:29,689][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864512 seconds
+[2025-12-18 14:49:29,704][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921570 seconds
+[2025-12-18 14:49:29,747][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797237 seconds
+[2025-12-18 14:49:29,778][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969818 seconds
+[2025-12-18 14:49:29,780][openai._base_client][INFO] - Retrying request to /chat/completions in 0.800991 seconds
+[2025-12-18 14:49:30,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:30,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:30,783][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:32,204][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:32,475][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:32,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:32,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:33,652][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:33,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:38,800][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416971 seconds
+[2025-12-18 14:49:43,811][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473626 seconds
+[2025-12-18 14:49:43,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457091 seconds
+[2025-12-18 14:49:43,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414438 seconds
+[2025-12-18 14:49:43,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453476 seconds
+[2025-12-18 14:49:44,222][openai._base_client][INFO] - Retrying request to /chat/completions in 0.811252 seconds
+[2025-12-18 14:49:46,760][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:49,234][openai._base_client][INFO] - Retrying request to /chat/completions in 0.853993 seconds
+[2025-12-18 14:49:49,270][openai._base_client][INFO] - Retrying request to /chat/completions in 0.966455 seconds
+[2025-12-18 14:49:49,273][openai._base_client][INFO] - Retrying request to /chat/completions in 0.890446 seconds
+[2025-12-18 14:49:52,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:49:53,446][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:00,176][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482988 seconds
+[2025-12-18 14:50:04,961][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:05,183][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455886 seconds
+[2025-12-18 14:50:05,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:05,664][openai._base_client][INFO] - Retrying request to /chat/completions in 0.975958 seconds
+[2025-12-18 14:50:09,589][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:11,565][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:16,574][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436067 seconds
+[2025-12-18 14:50:19,425][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:21,579][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493125 seconds
+[2025-12-18 14:50:22,014][openai._base_client][INFO] - Retrying request to /chat/completions in 0.779423 seconds
+[2025-12-18 14:50:24,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:32,808][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430530 seconds
+[2025-12-18 14:50:36,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:36,761][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:36,810][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:41,818][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497551 seconds
+[2025-12-18 14:50:45,784][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:47,473][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:52,482][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437441 seconds
+[2025-12-18 14:50:55,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:50:57,927][openai._base_client][INFO] - Retrying request to /chat/completions in 0.845600 seconds
+[2025-12-18 14:51:01,459][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:51:06,468][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462704 seconds
+[2025-12-18 14:51:11,939][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916366 seconds
+[2025-12-18 14:51:30,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:51:35,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424197 seconds
+[2025-12-18 14:51:36,306][openai._base_client][INFO] - Retrying request to /chat/completions in 0.758664 seconds
+[2025-12-18 14:51:42,169][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461415 seconds
+[2025-12-18 14:51:42,686][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905531 seconds
+[2025-12-18 14:51:48,769][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438218 seconds
+[2025-12-18 14:51:49,321][openai._base_client][INFO] - Retrying request to /chat/completions in 0.992759 seconds
+[2025-12-18 14:51:55,502][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477987 seconds
+[2025-12-18 14:51:56,096][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842154 seconds
+[2025-12-18 14:52:08,553][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:52:13,564][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421285 seconds
+[2025-12-18 14:52:14,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.784450 seconds
+[2025-12-18 14:52:29,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:52:34,559][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498527 seconds
+[2025-12-18 14:52:35,178][openai._base_client][INFO] - Retrying request to /chat/completions in 0.950328 seconds
+[2025-12-18 14:52:41,342][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449058 seconds
+[2025-12-18 14:52:41,849][openai._base_client][INFO] - Retrying request to /chat/completions in 0.965823 seconds
+[2025-12-18 14:52:47,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:52:52,177][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486808 seconds
+[2025-12-18 14:52:52,717][openai._base_client][INFO] - Retrying request to /chat/completions in 0.894722 seconds
+[2025-12-18 14:52:58,724][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420712 seconds
+[2025-12-18 14:53:03,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:03,642][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490530 seconds
+[2025-12-18 14:53:03,694][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475585 seconds
+[2025-12-18 14:53:03,696][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408910 seconds
+[2025-12-18 14:53:03,697][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450754 seconds
+[2025-12-18 14:53:03,754][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406203 seconds
+[2025-12-18 14:53:03,757][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471206 seconds
+[2025-12-18 14:53:03,763][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455997 seconds
+[2025-12-18 14:53:04,161][openai._base_client][INFO] - Retrying request to /chat/completions in 0.919989 seconds
+[2025-12-18 14:53:04,195][openai._base_client][INFO] - Retrying request to /chat/completions in 0.970080 seconds
+[2025-12-18 14:53:04,203][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979887 seconds
+[2025-12-18 14:53:04,223][openai._base_client][INFO] - Retrying request to /chat/completions in 0.907204 seconds
+[2025-12-18 14:53:04,225][openai._base_client][INFO] - Retrying request to /chat/completions in 0.941235 seconds
+[2025-12-18 14:53:04,270][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944544 seconds
+[2025-12-18 14:53:04,282][openai._base_client][INFO] - Retrying request to /chat/completions in 0.883305 seconds
+[2025-12-18 14:53:05,268][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492938 seconds
+[2025-12-18 14:53:05,270][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418595 seconds
+[2025-12-18 14:53:05,274][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452084 seconds
+[2025-12-18 14:53:05,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:05,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:05,743][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920040 seconds
+[2025-12-18 14:53:05,774][openai._base_client][INFO] - Retrying request to /chat/completions in 0.966608 seconds
+[2025-12-18 14:53:05,922][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:05,933][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:06,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:06,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:06,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:06,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483469 seconds
+[2025-12-18 14:53:06,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:07,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:07,337][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:07,711][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:07,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,109][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,123][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,588][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,645][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454360 seconds
+[2025-12-18 14:53:08,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:08,901][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393137 seconds
+[2025-12-18 14:53:09,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:09,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:09,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:10,472][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:10,714][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:11,583][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:16,594][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449864 seconds
+[2025-12-18 14:53:16,652][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420790 seconds
+[2025-12-18 14:53:17,096][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870966 seconds
+[2025-12-18 14:53:18,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:18,608][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:18,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:23,717][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402128 seconds
+[2025-12-18 14:53:26,285][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:26,386][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389302 seconds
+[2025-12-18 14:53:26,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399000 seconds
+[2025-12-18 14:53:26,434][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437743 seconds
+[2025-12-18 14:53:26,437][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386667 seconds
+[2025-12-18 14:53:26,438][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392468 seconds
+[2025-12-18 14:53:26,499][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482818 seconds
+[2025-12-18 14:53:26,500][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484807 seconds
+[2025-12-18 14:53:26,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453654 seconds
+[2025-12-18 14:53:26,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944003 seconds
+[2025-12-18 14:53:26,882][openai._base_client][INFO] - Retrying request to /chat/completions in 0.927741 seconds
+[2025-12-18 14:53:26,886][openai._base_client][INFO] - Retrying request to /chat/completions in 0.907912 seconds
+[2025-12-18 14:53:26,925][openai._base_client][INFO] - Retrying request to /chat/completions in 0.835493 seconds
+[2025-12-18 14:53:27,016][openai._base_client][INFO] - Retrying request to /chat/completions in 0.891330 seconds
+[2025-12-18 14:53:27,034][openai._base_client][INFO] - Retrying request to /chat/completions in 0.852676 seconds
+[2025-12-18 14:53:27,036][openai._base_client][INFO] - Retrying request to /chat/completions in 0.992595 seconds
+[2025-12-18 14:53:27,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:27,866][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480711 seconds
+[2025-12-18 14:53:27,878][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485537 seconds
+[2025-12-18 14:53:27,897][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466326 seconds
+[2025-12-18 14:53:27,912][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388667 seconds
+[2025-12-18 14:53:27,992][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396776 seconds
+[2025-12-18 14:53:28,019][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381778 seconds
+[2025-12-18 14:53:28,132][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480037 seconds
+[2025-12-18 14:53:28,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:28,365][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:28,399][openai._base_client][INFO] - Retrying request to /chat/completions in 0.988843 seconds
+[2025-12-18 14:53:28,413][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816562 seconds
+[2025-12-18 14:53:28,421][openai._base_client][INFO] - Retrying request to /chat/completions in 0.827945 seconds
+[2025-12-18 14:53:28,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:28,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.938357 seconds
+[2025-12-18 14:53:28,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:29,037][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:29,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:29,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:29,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:29,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:30,290][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:30,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:31,325][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:31,509][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:31,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:31,827][openai._base_client][INFO] - Retrying request to /chat/completions in 0.911729 seconds
+[2025-12-18 14:53:34,468][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:36,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:41,347][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424340 seconds
+[2025-12-18 14:53:41,471][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483491 seconds
+[2025-12-18 14:53:41,473][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407866 seconds
+[2025-12-18 14:53:41,827][openai._base_client][INFO] - Retrying request to /chat/completions in 0.987229 seconds
+[2025-12-18 14:53:41,939][openai._base_client][INFO] - Retrying request to /chat/completions in 0.868479 seconds
+[2025-12-18 14:53:42,012][openai._base_client][INFO] - Retrying request to /chat/completions in 0.853063 seconds
+[2025-12-18 14:53:42,665][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:42,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:43,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:45,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:45,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:46,363][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433013 seconds
+[2025-12-18 14:53:52,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:57,875][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403036 seconds
+[2025-12-18 14:53:59,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:53:59,567][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391998 seconds
+[2025-12-18 14:54:00,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:01,106][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:01,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:01,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:01,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:01,512][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:01,882][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:02,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:02,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:04,215][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:04,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:04,473][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:04,485][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:04,569][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400023 seconds
+[2025-12-18 14:54:04,574][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429851 seconds
+[2025-12-18 14:54:04,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:04,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:04,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:05,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:05,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:05,971][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471455 seconds
+[2025-12-18 14:54:05,973][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389983 seconds
+[2025-12-18 14:54:05,975][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435752 seconds
+[2025-12-18 14:54:05,977][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410074 seconds
+[2025-12-18 14:54:05,979][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408143 seconds
+[2025-12-18 14:54:05,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455569 seconds
+[2025-12-18 14:54:05,982][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466145 seconds
+[2025-12-18 14:54:05,983][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489330 seconds
+[2025-12-18 14:54:05,984][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389042 seconds
+[2025-12-18 14:54:05,984][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439770 seconds
+[2025-12-18 14:54:05,985][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391408 seconds
+[2025-12-18 14:54:05,985][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462006 seconds
+[2025-12-18 14:54:06,366][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870037 seconds
+[2025-12-18 14:54:06,376][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815395 seconds
+[2025-12-18 14:54:07,757][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:07,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:07,997][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:08,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:09,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:09,696][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:19,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:19,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:19,303][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:19,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:19,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:19,776][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:19,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496545 seconds
+[2025-12-18 14:54:19,820][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420821 seconds
+[2025-12-18 14:54:19,822][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437235 seconds
+[2025-12-18 14:54:19,823][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479077 seconds
+[2025-12-18 14:54:20,245][openai._base_client][INFO] - Retrying request to /chat/completions in 0.892543 seconds
+[2025-12-18 14:54:20,262][openai._base_client][INFO] - Retrying request to /chat/completions in 0.928930 seconds
+[2025-12-18 14:54:22,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:23,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:23,988][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:25,430][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:25,449][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382323 seconds
+[2025-12-18 14:54:25,834][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870995 seconds
+[2025-12-18 14:54:31,716][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400899 seconds
+[2025-12-18 14:54:35,468][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:54:35,487][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445706 seconds
+[2025-12-18 14:54:37,784][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/14-57-52/.hydra/config.yaml b/outputs/2025-12-18/14-57-52/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d28de1e7fab758543e07784afa776dc2e07f1a2c
--- /dev/null
+++ b/outputs/2025-12-18/14-57-52/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/14-57-52/.hydra/hydra.yaml b/outputs/2025-12-18/14-57-52/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bf983c78c94ca10df6204067745dd47fc88e3e44
--- /dev/null
+++ b/outputs/2025-12-18/14-57-52/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/14-57-52
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/14-57-52/.hydra/overrides.yaml b/outputs/2025-12-18/14-57-52/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/14-57-52/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/14-57-52/eval_api.log b/outputs/2025-12-18/14-57-52/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..f75fcf33a8773e9e51796de1d6910990a450497d
--- /dev/null
+++ b/outputs/2025-12-18/14-57-52/eval_api.log
@@ -0,0 +1,122 @@
+[2025-12-18 14:57:58,230][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:58,337][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:58,360][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:58,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:58,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:58,558][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:59,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:59,287][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:59,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:57:59,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:00,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:00,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:00,411][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418662 seconds
+[2025-12-18 14:58:00,411][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392354 seconds
+[2025-12-18 14:58:00,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:00,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:00,577][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:00,972][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:01,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:01,587][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:01,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:01,993][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:02,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:02,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:02,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:02,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:02,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:03,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:03,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:03,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:03,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:03,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:03,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:05,808][openai._base_client][INFO] - Retrying request to /chat/completions in 0.820057 seconds
+[2025-12-18 14:58:08,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:13,353][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477234 seconds
+[2025-12-18 14:58:17,534][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:18,835][openai._base_client][INFO] - Retrying request to /chat/completions in 0.827571 seconds
+[2025-12-18 14:58:18,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:22,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:27,736][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386946 seconds
+[2025-12-18 14:58:32,746][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452999 seconds
+[2025-12-18 14:58:32,748][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499493 seconds
+[2025-12-18 14:58:33,127][openai._base_client][INFO] - Retrying request to /chat/completions in 0.820374 seconds
+[2025-12-18 14:58:35,931][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:37,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:38,205][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774239 seconds
+[2025-12-18 14:58:41,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:46,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399335 seconds
+[2025-12-18 14:58:51,636][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399754 seconds
+[2025-12-18 14:58:52,033][openai._base_client][INFO] - Retrying request to /chat/completions in 0.874458 seconds
+[2025-12-18 14:58:55,020][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:58:55,335][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:59:00,343][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435175 seconds
+[2025-12-18 14:59:05,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803121 seconds
+[2025-12-18 14:59:08,827][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:59:13,836][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456458 seconds
+[2025-12-18 14:59:19,301][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985644 seconds
+[2025-12-18 14:59:35,309][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379400 seconds
+[2025-12-18 14:59:40,029][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 14:59:45,039][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377504 seconds
+[2025-12-18 14:59:50,425][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806185 seconds
+[2025-12-18 15:00:06,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:00:11,446][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414010 seconds
+[2025-12-18 15:00:15,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:00:20,506][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390019 seconds
+[2025-12-18 15:00:23,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:00:28,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485670 seconds
+[2025-12-18 15:00:35,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:00:40,103][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489103 seconds
+[2025-12-18 15:00:46,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:00:51,283][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436600 seconds
+[2025-12-18 15:00:58,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:01:03,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498924 seconds
+[2025-12-18 15:01:09,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:01:14,147][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476358 seconds
+[2025-12-18 15:01:18,476][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:01:23,486][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389101 seconds
+[2025-12-18 15:01:27,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:01:32,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491959 seconds
+[2025-12-18 15:01:37,757][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812345 seconds
+[2025-12-18 15:01:43,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:01:48,515][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391395 seconds
+[2025-12-18 15:02:03,648][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:02:08,658][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411564 seconds
+[2025-12-18 15:02:13,073][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:02:18,085][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387092 seconds
+[2025-12-18 15:02:22,428][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:02:27,444][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462465 seconds
+[2025-12-18 15:02:32,915][openai._base_client][INFO] - Retrying request to /chat/completions in 0.923900 seconds
+[2025-12-18 15:02:37,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:02:42,841][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408848 seconds
+[2025-12-18 15:02:47,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:02:52,075][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393241 seconds
+[2025-12-18 15:02:57,478][openai._base_client][INFO] - Retrying request to /chat/completions in 0.769308 seconds
+[2025-12-18 15:03:01,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:03:06,206][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496830 seconds
+[2025-12-18 15:03:09,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:03:14,093][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498728 seconds
+[2025-12-18 15:03:17,164][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:03:22,174][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416638 seconds
+[2025-12-18 15:03:26,756][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:03:31,765][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432357 seconds
+[2025-12-18 15:03:36,415][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:03:41,425][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449097 seconds
+[2025-12-18 15:03:47,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:03:52,033][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390485 seconds
+[2025-12-18 15:03:57,432][openai._base_client][INFO] - Retrying request to /chat/completions in 0.938874 seconds
+[2025-12-18 15:04:05,997][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:04:11,007][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387111 seconds
+[2025-12-18 15:04:16,403][openai._base_client][INFO] - Retrying request to /chat/completions in 0.947921 seconds
+[2025-12-18 15:04:21,741][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:04:26,750][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376439 seconds
+[2025-12-18 15:04:32,136][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797820 seconds
+[2025-12-18 15:04:35,909][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:04:40,920][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487790 seconds
+[2025-12-18 15:04:46,417][openai._base_client][INFO] - Retrying request to /chat/completions in 0.912195 seconds
+[2025-12-18 15:04:50,013][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:04:56,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480463 seconds
+[2025-12-18 15:05:04,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:05:09,218][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417171 seconds
+[2025-12-18 15:05:12,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-05-38/.hydra/hydra.yaml b/outputs/2025-12-18/15-05-38/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dbb7c8616aa24e5854abbb020d76e480398c158d
--- /dev/null
+++ b/outputs/2025-12-18/15-05-38/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-05-38
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-05-38/.hydra/overrides.yaml b/outputs/2025-12-18/15-05-38/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-05-38/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-05-38/eval_api.log b/outputs/2025-12-18/15-05-38/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..19663ab3623aa7d63ac92c3cbaa1395c5e7938a8
--- /dev/null
+++ b/outputs/2025-12-18/15-05-38/eval_api.log
@@ -0,0 +1,75 @@
+[2025-12-18 15:05:49,101][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7fbcc5387970>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 721f7313-14a5-4752-ac6f-30dfde4363b6)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 15:05:49,103][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 15:06:00,397][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /api/resolve-cache/models/Qwen/Qwen2.5-0.5B-Instruct/7ae557604adf67be50417f59c2c2f167def9a775/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7fbcc53aab60>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 737fd684-d003-47f3-ac15-b90851cb0bf5)')' thrown while requesting HEAD https://huggingface.co/api/resolve-cache/models/Qwen/Qwen2.5-0.5B-Instruct/7ae557604adf67be50417f59c2c2f167def9a775/tokenizer_config.json
+[2025-12-18 15:06:00,399][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 15:06:06,429][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:06,881][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:07,323][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:07,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:07,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:07,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:08,312][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390809 seconds
+[2025-12-18 15:06:08,313][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473352 seconds
+[2025-12-18 15:06:08,313][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438027 seconds
+[2025-12-18 15:06:08,315][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427177 seconds
+[2025-12-18 15:06:08,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:08,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:08,544][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:08,695][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:08,890][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:09,283][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:09,519][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:09,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:10,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:10,641][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:11,470][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:11,589][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:11,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:11,695][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:11,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:11,838][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:11,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:12,197][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:12,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:12,652][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:12,957][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:13,095][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:13,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:13,888][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:14,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:15,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:20,638][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440947 seconds
+[2025-12-18 15:06:25,644][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499008 seconds
+[2025-12-18 15:06:25,729][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:27,962][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:30,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:35,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432070 seconds
+[2025-12-18 15:06:40,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:41,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.906567 seconds
+[2025-12-18 15:06:45,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:50,560][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376160 seconds
+[2025-12-18 15:06:54,029][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:54,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:06:59,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379920 seconds
+[2025-12-18 15:07:04,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:07:09,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461336 seconds
+[2025-12-18 15:07:12,530][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:07:17,540][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472645 seconds
+[2025-12-18 15:07:23,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969281 seconds
+[2025-12-18 15:07:39,012][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409120 seconds
+[2025-12-18 15:07:46,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:07:51,716][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391829 seconds
+[2025-12-18 15:07:58,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:08:03,986][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488296 seconds
+[2025-12-18 15:08:10,477][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:08:15,486][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485921 seconds
+[2025-12-18 15:08:21,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:08:26,846][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457565 seconds
+[2025-12-18 15:08:33,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:08:38,421][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497153 seconds
+[2025-12-18 15:08:44,133][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:08:49,142][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497727 seconds
+[2025-12-18 15:08:54,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:08:59,228][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413196 seconds
+[2025-12-18 15:09:04,180][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:09:09,189][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415572 seconds
diff --git a/outputs/2025-12-18/15-10-02/.hydra/config.yaml b/outputs/2025-12-18/15-10-02/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..59425ef114885b47f6e1909be3a3f2e7a02c79d0
--- /dev/null
+++ b/outputs/2025-12-18/15-10-02/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-10-02/.hydra/hydra.yaml b/outputs/2025-12-18/15-10-02/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..facbbf62ec3593658774fdda61d4a7d4c89566ef
--- /dev/null
+++ b/outputs/2025-12-18/15-10-02/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-10-02
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-10-02/.hydra/overrides.yaml b/outputs/2025-12-18/15-10-02/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-10-02/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-10-02/eval_api.log b/outputs/2025-12-18/15-10-02/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..86c02a536c305b2a4d3039434e7e82a669ca54dc
--- /dev/null
+++ b/outputs/2025-12-18/15-10-02/eval_api.log
@@ -0,0 +1,23 @@
+[2025-12-18 15:10:12,541][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f7811982680>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 2b00a14a-bbbd-42de-8b82-004a73f16874)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 15:10:12,543][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 15:10:23,852][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /api/resolve-cache/models/Qwen/Qwen2.5-0.5B-Instruct/7ae557604adf67be50417f59c2c2f167def9a775/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f7811998850>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 6b5e8a33-fcc0-4d4d-8db1-4e5f169bc90e)')' thrown while requesting HEAD https://huggingface.co/api/resolve-cache/models/Qwen/Qwen2.5-0.5B-Instruct/7ae557604adf67be50417f59c2c2f167def9a775/tokenizer_config.json
+[2025-12-18 15:10:23,853][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 15:10:30,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:10:30,644][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494158 seconds
+[2025-12-18 15:10:34,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:10:39,886][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437334 seconds
+[2025-12-18 15:10:45,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.855650 seconds
+[2025-12-18 15:10:49,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:10:54,862][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385398 seconds
+[2025-12-18 15:11:00,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.911245 seconds
+[2025-12-18 15:11:03,631][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:11:08,641][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377245 seconds
+[2025-12-18 15:11:14,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.894415 seconds
+[2025-12-18 15:11:27,159][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:11:27,169][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437674 seconds
+[2025-12-18 15:11:32,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.976376 seconds
+[2025-12-18 15:11:35,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:11:35,341][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442323 seconds
+[2025-12-18 15:11:39,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:11:39,168][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434758 seconds
+[2025-12-18 15:11:40,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-17-02/.hydra/config.yaml b/outputs/2025-12-18/15-17-02/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cbbeb404106c8d574cdeb37d35d4a84e6762d9c7
--- /dev/null
+++ b/outputs/2025-12-18/15-17-02/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-17-02/.hydra/hydra.yaml b/outputs/2025-12-18/15-17-02/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a4e39c4d427d23c60e20e48740c22d4e58fd3631
--- /dev/null
+++ b/outputs/2025-12-18/15-17-02/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-17-02
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-17-02/.hydra/overrides.yaml b/outputs/2025-12-18/15-17-02/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-17-02/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-17-02/eval_api.log b/outputs/2025-12-18/15-17-02/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..651f40ad784f900a36d85321777912e680e4065c
--- /dev/null
+++ b/outputs/2025-12-18/15-17-02/eval_api.log
@@ -0,0 +1,171 @@
+[2025-12-18 15:17:05,140][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406007 seconds
+[2025-12-18 15:17:05,151][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457064 seconds
+[2025-12-18 15:17:08,475][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:08,694][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:08,928][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:09,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:09,648][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:10,004][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390810 seconds
+[2025-12-18 15:17:10,005][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398033 seconds
+[2025-12-18 15:17:10,006][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387260 seconds
+[2025-12-18 15:17:10,033][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:10,557][openai._base_client][INFO] - Retrying request to /chat/completions in 0.758886 seconds
+[2025-12-18 15:17:10,949][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:11,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:12,408][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:12,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:13,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:14,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:14,468][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:16,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:16,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:16,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:21,551][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427533 seconds
+[2025-12-18 15:17:25,885][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:26,983][openai._base_client][INFO] - Retrying request to /chat/completions in 0.826199 seconds
+[2025-12-18 15:17:27,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:32,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:32,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423770 seconds
+[2025-12-18 15:17:32,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424689 seconds
+[2025-12-18 15:17:32,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495473 seconds
+[2025-12-18 15:17:32,585][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436987 seconds
+[2025-12-18 15:17:32,586][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463223 seconds
+[2025-12-18 15:17:32,587][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406500 seconds
+[2025-12-18 15:17:32,588][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411198 seconds
+[2025-12-18 15:17:33,007][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901808 seconds
+[2025-12-18 15:17:33,050][openai._base_client][INFO] - Retrying request to /chat/completions in 0.977472 seconds
+[2025-12-18 15:17:33,076][openai._base_client][INFO] - Retrying request to /chat/completions in 0.836447 seconds
+[2025-12-18 15:17:33,101][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818628 seconds
+[2025-12-18 15:17:33,129][openai._base_client][INFO] - Retrying request to /chat/completions in 0.809284 seconds
+[2025-12-18 15:17:34,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:34,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:34,868][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:35,595][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:36,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:37,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:37,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490395 seconds
+[2025-12-18 15:17:37,537][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408850 seconds
+[2025-12-18 15:17:37,537][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446924 seconds
+[2025-12-18 15:17:37,539][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473309 seconds
+[2025-12-18 15:17:39,076][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:39,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:17:40,652][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:01,858][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:06,868][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455687 seconds
+[2025-12-18 15:18:07,377][openai._base_client][INFO] - Retrying request to /chat/completions in 0.787624 seconds
+[2025-12-18 15:18:09,511][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:09,693][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:11,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:16,428][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469451 seconds
+[2025-12-18 15:18:18,751][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:18,814][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384880 seconds
+[2025-12-18 15:18:18,862][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424862 seconds
+[2025-12-18 15:18:18,867][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491061 seconds
+[2025-12-18 15:18:18,869][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482302 seconds
+[2025-12-18 15:18:18,869][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430741 seconds
+[2025-12-18 15:18:18,870][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411623 seconds
+[2025-12-18 15:18:18,872][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417760 seconds
+[2025-12-18 15:18:18,872][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413487 seconds
+[2025-12-18 15:18:19,338][openai._base_client][INFO] - Retrying request to /chat/completions in 0.875360 seconds
+[2025-12-18 15:18:19,340][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939379 seconds
+[2025-12-18 15:18:19,407][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842564 seconds
+[2025-12-18 15:18:21,046][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:22,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:22,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:22,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:22,741][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:22,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:23,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:23,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:23,500][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:23,534][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:23,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:26,316][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:26,380][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475814 seconds
+[2025-12-18 15:18:26,382][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414334 seconds
+[2025-12-18 15:18:26,386][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402924 seconds
+[2025-12-18 15:18:26,387][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492838 seconds
+[2025-12-18 15:18:26,389][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400452 seconds
+[2025-12-18 15:18:26,391][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494502 seconds
+[2025-12-18 15:18:26,392][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495290 seconds
+[2025-12-18 15:18:26,394][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395496 seconds
+[2025-12-18 15:18:26,397][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375262 seconds
+[2025-12-18 15:18:26,398][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400798 seconds
+[2025-12-18 15:18:26,775][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844491 seconds
+[2025-12-18 15:18:26,844][openai._base_client][INFO] - Retrying request to /chat/completions in 0.795915 seconds
+[2025-12-18 15:18:26,845][openai._base_client][INFO] - Retrying request to /chat/completions in 0.919494 seconds
+[2025-12-18 15:18:28,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:28,887][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:29,283][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:29,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:30,154][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:30,522][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:30,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:31,219][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:31,795][openai._base_client][INFO] - Retrying request to /chat/completions in 0.872538 seconds
+[2025-12-18 15:18:34,072][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:39,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478938 seconds
+[2025-12-18 15:18:39,613][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921518 seconds
+[2025-12-18 15:18:42,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:42,744][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498618 seconds
+[2025-12-18 15:18:44,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:44,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:44,878][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:45,128][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:45,166][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:46,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:46,913][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:46,955][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497331 seconds
+[2025-12-18 15:18:46,956][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420228 seconds
+[2025-12-18 15:18:46,957][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432543 seconds
+[2025-12-18 15:18:46,958][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379331 seconds
+[2025-12-18 15:18:46,958][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420349 seconds
+[2025-12-18 15:18:47,341][openai._base_client][INFO] - Retrying request to /chat/completions in 0.886902 seconds
+[2025-12-18 15:18:47,381][openai._base_client][INFO] - Retrying request to /chat/completions in 0.898907 seconds
+[2025-12-18 15:18:47,427][openai._base_client][INFO] - Retrying request to /chat/completions in 0.820393 seconds
+[2025-12-18 15:18:49,847][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:49,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:50,212][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:54,413][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:18:59,423][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389631 seconds
+[2025-12-18 15:19:01,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:03,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:03,549][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492866 seconds
+[2025-12-18 15:19:03,552][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414440 seconds
+[2025-12-18 15:19:03,604][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437869 seconds
+[2025-12-18 15:19:04,093][openai._base_client][INFO] - Retrying request to /chat/completions in 0.750702 seconds
+[2025-12-18 15:19:05,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:06,002][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:06,182][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:06,953][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:07,058][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:07,085][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495911 seconds
+[2025-12-18 15:19:07,086][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379364 seconds
+[2025-12-18 15:19:07,468][openai._base_client][INFO] - Retrying request to /chat/completions in 0.881433 seconds
+[2025-12-18 15:19:07,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815615 seconds
+[2025-12-18 15:19:13,526][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429617 seconds
+[2025-12-18 15:19:18,474][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425576 seconds
+[2025-12-18 15:19:18,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:20,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:25,411][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420525 seconds
+[2025-12-18 15:19:25,884][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816625 seconds
+[2025-12-18 15:19:31,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492731 seconds
+[2025-12-18 15:19:40,258][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:45,270][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490480 seconds
+[2025-12-18 15:19:49,850][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:49,881][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498033 seconds
+[2025-12-18 15:19:51,877][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:19:54,884][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483901 seconds
+[2025-12-18 15:19:58,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:20:03,923][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446972 seconds
+[2025-12-18 15:20:09,377][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803030 seconds
+[2025-12-18 15:20:20,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:20:25,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412168 seconds
+[2025-12-18 15:20:30,943][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:20:35,954][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389566 seconds
+[2025-12-18 15:20:36,401][openai._base_client][INFO] - Retrying request to /chat/completions in 0.829523 seconds
+[2025-12-18 15:20:47,299][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376479 seconds
+[2025-12-18 15:20:50,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:20:50,884][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457481 seconds
+[2025-12-18 15:20:56,350][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955798 seconds
+[2025-12-18 15:20:59,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-21-45/.hydra/config.yaml b/outputs/2025-12-18/15-21-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8d56bfbc285ed6b97bbc82c39f35bed5f99a145c
--- /dev/null
+++ b/outputs/2025-12-18/15-21-45/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-21-45/.hydra/hydra.yaml b/outputs/2025-12-18/15-21-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..72a3410be020a067e4c35784a785af8bf9420ad4
--- /dev/null
+++ b/outputs/2025-12-18/15-21-45/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-21-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-21-45/.hydra/overrides.yaml b/outputs/2025-12-18/15-21-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-21-45/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-21-45/eval_api.log b/outputs/2025-12-18/15-21-45/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..9b2951d52d0374826ad64a5524b7661fe2683712
--- /dev/null
+++ b/outputs/2025-12-18/15-21-45/eval_api.log
@@ -0,0 +1,159 @@
+[2025-12-18 15:21:50,745][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:50,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:50,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:50,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:51,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:51,270][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:51,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:51,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:52,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:52,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:52,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:52,694][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:52,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:52,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:52,980][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424859 seconds
+[2025-12-18 15:21:52,982][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427182 seconds
+[2025-12-18 15:21:53,054][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:53,260][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:54,024][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:54,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:54,437][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:54,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:54,831][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:55,024][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:55,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:55,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:55,770][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:55,789][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:56,342][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:56,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:56,840][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:56,918][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:57,091][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:57,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:57,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:57,519][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:58,072][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:58,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:58,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:58,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:58,412][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863954 seconds
+[2025-12-18 15:21:58,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:59,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:59,253][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:59,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:59,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:59,745][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:21:59,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:00,076][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:00,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:00,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:00,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:01,282][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:01,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:01,596][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:01,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:02,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:02,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,085][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,328][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,620][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,666][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:03,776][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:04,392][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:04,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:04,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:05,123][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:05,275][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:05,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:05,544][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:05,843][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:05,883][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:06,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:06,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:06,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:06,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:06,602][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:07,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:07,214][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:07,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:07,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:07,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:07,809][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:08,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:08,851][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:09,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:09,172][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:09,283][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479440 seconds
+[2025-12-18 15:22:09,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:09,389][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:09,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:10,388][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:10,396][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:10,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:10,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:10,845][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:10,881][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:11,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:11,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:11,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:11,229][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:11,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:11,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:11,626][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,448][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,584][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,609][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,775][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:12,882][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:13,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:13,229][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:13,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:13,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:13,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:14,338][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:14,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:14,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:14,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:14,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:15,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:15,206][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:17,132][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:22,143][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463386 seconds
+[2025-12-18 15:22:25,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:25,732][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:25,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:25,988][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:26,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:26,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:27,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:27,612][openai._base_client][INFO] - Retrying request to /chat/completions in 0.782598 seconds
+[2025-12-18 15:22:27,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:30,392][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:35,403][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481681 seconds
+[2025-12-18 15:22:39,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:40,411][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381015 seconds
+[2025-12-18 15:22:40,413][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390543 seconds
+[2025-12-18 15:22:40,413][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419302 seconds
+[2025-12-18 15:22:43,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:44,265][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:45,807][openai._base_client][INFO] - Retrying request to /chat/completions in 0.961160 seconds
+[2025-12-18 15:22:49,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:54,455][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497098 seconds
+[2025-12-18 15:22:56,722][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:57,525][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:22:59,463][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446456 seconds
+[2025-12-18 15:23:02,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:07,510][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399456 seconds
+[2025-12-18 15:23:09,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:11,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-23-26/.hydra/config.yaml b/outputs/2025-12-18/15-23-26/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3d6d8740208ad87a5eb30c32278a18a42a2ae39f
--- /dev/null
+++ b/outputs/2025-12-18/15-23-26/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-23-26/.hydra/hydra.yaml b/outputs/2025-12-18/15-23-26/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..251ae8ef1bb00d542f3c428715e170f06e806ef2
--- /dev/null
+++ b/outputs/2025-12-18/15-23-26/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-23-26
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-23-26/.hydra/overrides.yaml b/outputs/2025-12-18/15-23-26/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-23-26/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-23-26/eval_api.log b/outputs/2025-12-18/15-23-26/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..72314927af067d9e9d5f53542df5d7de855eb8a9
--- /dev/null
+++ b/outputs/2025-12-18/15-23-26/eval_api.log
@@ -0,0 +1,960 @@
+[2025-12-18 15:23:36,995][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7fb649e79420>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: dcf67d70-365a-45fa-9006-4534e24c86ac)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 15:23:36,997][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 15:23:42,989][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,127][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,478][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:43,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:44,992][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:45,011][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:45,107][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:45,177][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:45,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:45,274][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378756 seconds
+[2025-12-18 15:23:45,430][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:45,584][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:46,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:46,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:46,909][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:46,945][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:46,981][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:47,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:47,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:47,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:47,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:47,498][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:48,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:48,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:48,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,363][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,378][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,533][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,717][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:49,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:50,197][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:50,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:50,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:50,658][openai._base_client][INFO] - Retrying request to /chat/completions in 0.771669 seconds
+[2025-12-18 15:23:51,209][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:51,271][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:51,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:51,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:51,794][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:51,901][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:51,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,132][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,136][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,182][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,208][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,260][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,495][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:52,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:53,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,433][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,525][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,589][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,592][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:54,689][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:55,275][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:55,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:55,405][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:55,489][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:55,494][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:56,317][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:56,452][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:56,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:56,644][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:56,717][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:56,771][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:56,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:57,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:57,192][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:57,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:57,619][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:57,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:57,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:57,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:58,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:58,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:58,976][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,001][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,103][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,159][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,256][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:23:59,913][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:00,244][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:00,553][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:00,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:00,942][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:01,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:01,434][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:01,637][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:01,932][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,470][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:02,677][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:03,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:03,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,047][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,148][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,250][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,464][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,568][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:04,869][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:05,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:10,031][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416924 seconds
+[2025-12-18 15:24:10,481][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448483 seconds
+[2025-12-18 15:24:12,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:12,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:12,968][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,177][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,310][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,416][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,525][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,827][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:13,992][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:14,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,049][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427642 seconds
+[2025-12-18 15:24:15,050][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411485 seconds
+[2025-12-18 15:24:15,050][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437116 seconds
+[2025-12-18 15:24:15,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,831][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870944 seconds
+[2025-12-18 15:24:15,835][openai._base_client][INFO] - Retrying request to /chat/completions in 0.761906 seconds
+[2025-12-18 15:24:15,840][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,950][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,981][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:15,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:16,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:16,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:16,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:17,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:18,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:18,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:18,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:18,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:18,624][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:18,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:18,962][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:19,159][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:19,437][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:19,494][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:19,530][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:19,850][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:19,989][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:20,170][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:20,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:20,817][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:20,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:20,878][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:21,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:21,109][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:21,175][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:21,339][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:21,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:21,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:22,038][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:22,214][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:22,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:22,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:22,775][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:23,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:23,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:23,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:23,798][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:23,833][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:24,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:24,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:24,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:24,266][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:24,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:24,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:25,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:25,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:25,250][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:25,523][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:25,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:26,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:26,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:26,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:26,658][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:26,812][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:26,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:26,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:27,060][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:27,175][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:27,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:27,925][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:28,022][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:28,287][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:28,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:28,843][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,058][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,470][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:29,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:30,148][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:30,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:30,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:30,559][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:30,648][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:35,658][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489184 seconds
+[2025-12-18 15:24:36,086][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454688 seconds
+[2025-12-18 15:24:36,086][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484693 seconds
+[2025-12-18 15:24:36,088][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444429 seconds
+[2025-12-18 15:24:36,503][openai._base_client][INFO] - Retrying request to /chat/completions in 0.947750 seconds
+[2025-12-18 15:24:36,889][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863532 seconds
+[2025-12-18 15:24:36,929][openai._base_client][INFO] - Retrying request to /chat/completions in 0.935699 seconds
+[2025-12-18 15:24:38,154][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468346 seconds
+[2025-12-18 15:24:38,460][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393972 seconds
+[2025-12-18 15:24:38,928][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:38,973][openai._base_client][INFO] - Retrying request to /chat/completions in 0.978183 seconds
+[2025-12-18 15:24:39,079][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,090][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,209][openai._base_client][INFO] - Retrying request to /chat/completions in 0.970892 seconds
+[2025-12-18 15:24:39,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,459][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,488][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,577][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:39,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:40,675][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430258 seconds
+[2025-12-18 15:24:40,932][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499887 seconds
+[2025-12-18 15:24:41,459][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783068 seconds
+[2025-12-18 15:24:41,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.757334 seconds
+[2025-12-18 15:24:41,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:41,676][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:41,785][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:41,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:41,978][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:42,037][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:42,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:42,264][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:42,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:43,219][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419361 seconds
+[2025-12-18 15:24:43,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:44,163][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:44,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:44,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:44,664][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:44,672][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:44,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:44,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:45,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:45,310][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407708 seconds
+[2025-12-18 15:24:45,539][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489175 seconds
+[2025-12-18 15:24:45,667][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:46,381][openai._base_client][INFO] - Retrying request to /chat/completions in 0.872233 seconds
+[2025-12-18 15:24:46,436][openai._base_client][INFO] - Retrying request to /chat/completions in 0.776340 seconds
+[2025-12-18 15:24:46,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:47,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:47,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:47,597][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404658 seconds
+[2025-12-18 15:24:47,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:47,670][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470826 seconds
+[2025-12-18 15:24:47,915][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:47,955][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383127 seconds
+[2025-12-18 15:24:48,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:48,355][openai._base_client][INFO] - Retrying request to /chat/completions in 0.930506 seconds
+[2025-12-18 15:24:48,643][openai._base_client][INFO] - Retrying request to /chat/completions in 0.919583 seconds
+[2025-12-18 15:24:48,692][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944257 seconds
+[2025-12-18 15:24:48,930][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:49,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:49,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:49,659][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:50,020][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:50,273][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462715 seconds
+[2025-12-18 15:24:50,342][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:50,523][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:50,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:50,722][openai._base_client][INFO] - Retrying request to /chat/completions in 0.990105 seconds
+[2025-12-18 15:24:51,048][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:51,160][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:51,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:52,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:52,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:52,568][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406545 seconds
+[2025-12-18 15:24:52,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:52,828][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:52,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:53,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:53,153][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812903 seconds
+[2025-12-18 15:24:53,742][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:53,816][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:54,037][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:54,285][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:54,594][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:54,637][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433371 seconds
+[2025-12-18 15:24:54,668][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:54,995][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474246 seconds
+[2025-12-18 15:24:55,229][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:55,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:55,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:55,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:55,740][openai._base_client][INFO] - Retrying request to /chat/completions in 0.817697 seconds
+[2025-12-18 15:24:56,059][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:56,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:56,687][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:56,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,815][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,848][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,870][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:24:57,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.872946 seconds
+[2025-12-18 15:24:58,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:00,075][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904067 seconds
+[2025-12-18 15:25:00,473][openai._base_client][INFO] - Retrying request to /chat/completions in 0.843480 seconds
+[2025-12-18 15:25:01,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:03,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:04,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:09,441][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416005 seconds
+[2025-12-18 15:25:09,811][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432702 seconds
+[2025-12-18 15:25:09,816][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413068 seconds
+[2025-12-18 15:25:09,816][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425308 seconds
+[2025-12-18 15:25:09,817][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439857 seconds
+[2025-12-18 15:25:09,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391712 seconds
+[2025-12-18 15:25:09,820][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391248 seconds
+[2025-12-18 15:25:10,591][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878712 seconds
+[2025-12-18 15:25:10,595][openai._base_client][INFO] - Retrying request to /chat/completions in 0.938747 seconds
+[2025-12-18 15:25:12,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:12,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:13,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:14,462][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487834 seconds
+[2025-12-18 15:25:14,462][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412272 seconds
+[2025-12-18 15:25:14,464][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428748 seconds
+[2025-12-18 15:25:14,464][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387117 seconds
+[2025-12-18 15:25:14,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420302 seconds
+[2025-12-18 15:25:14,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376027 seconds
+[2025-12-18 15:25:14,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380009 seconds
+[2025-12-18 15:25:14,809][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:14,861][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999397 seconds
+[2025-12-18 15:25:14,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:15,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:15,214][openai._base_client][INFO] - Retrying request to /chat/completions in 0.882078 seconds
+[2025-12-18 15:25:15,228][openai._base_client][INFO] - Retrying request to /chat/completions in 0.961021 seconds
+[2025-12-18 15:25:15,232][openai._base_client][INFO] - Retrying request to /chat/completions in 0.827172 seconds
+[2025-12-18 15:25:15,261][openai._base_client][INFO] - Retrying request to /chat/completions in 0.805342 seconds
+[2025-12-18 15:25:15,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:16,826][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472556 seconds
+[2025-12-18 15:25:16,892][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447343 seconds
+[2025-12-18 15:25:17,312][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:17,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:17,477][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:17,653][openai._base_client][INFO] - Retrying request to /chat/completions in 0.980581 seconds
+[2025-12-18 15:25:17,695][openai._base_client][INFO] - Retrying request to /chat/completions in 0.884140 seconds
+[2025-12-18 15:25:17,771][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:18,008][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:18,126][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:18,704][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:19,284][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485395 seconds
+[2025-12-18 15:25:19,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:19,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:19,884][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:19,887][openai._base_client][INFO] - Retrying request to /chat/completions in 0.775388 seconds
+[2025-12-18 15:25:19,930][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:19,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.930480 seconds
+[2025-12-18 15:25:20,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:20,482][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:20,490][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:21,212][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379277 seconds
+[2025-12-18 15:25:21,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:21,416][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448495 seconds
+[2025-12-18 15:25:21,455][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422363 seconds
+[2025-12-18 15:25:22,224][openai._base_client][INFO] - Retrying request to /chat/completions in 0.794443 seconds
+[2025-12-18 15:25:22,468][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:22,561][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:22,667][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:22,768][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:23,024][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:23,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:23,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:23,737][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391146 seconds
+[2025-12-18 15:25:23,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:23,991][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379085 seconds
+[2025-12-18 15:25:24,485][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755026 seconds
+[2025-12-18 15:25:24,773][openai._base_client][INFO] - Retrying request to /chat/completions in 0.998215 seconds
+[2025-12-18 15:25:25,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:25,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:25,095][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:25,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:25,434][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:25,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:25,928][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:26,067][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395785 seconds
+[2025-12-18 15:25:26,236][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:26,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488242 seconds
+[2025-12-18 15:25:26,480][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490271 seconds
+[2025-12-18 15:25:26,595][openai._base_client][INFO] - Retrying request to /chat/completions in 0.981913 seconds
+[2025-12-18 15:25:26,821][openai._base_client][INFO] - Retrying request to /chat/completions in 0.888335 seconds
+[2025-12-18 15:25:26,882][openai._base_client][INFO] - Retrying request to /chat/completions in 0.964541 seconds
+[2025-12-18 15:25:27,390][openai._base_client][INFO] - Retrying request to /chat/completions in 0.896618 seconds
+[2025-12-18 15:25:27,885][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:27,952][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:28,278][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382461 seconds
+[2025-12-18 15:25:28,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:28,413][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437135 seconds
+[2025-12-18 15:25:28,626][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463947 seconds
+[2025-12-18 15:25:28,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:28,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:29,017][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774073 seconds
+[2025-12-18 15:25:29,098][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:29,200][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870338 seconds
+[2025-12-18 15:25:29,219][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470072 seconds
+[2025-12-18 15:25:29,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:29,374][openai._base_client][INFO] - Retrying request to /chat/completions in 0.751896 seconds
+[2025-12-18 15:25:29,443][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878163 seconds
+[2025-12-18 15:25:29,597][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:30,037][openai._base_client][INFO] - Retrying request to /chat/completions in 0.937743 seconds
+[2025-12-18 15:25:30,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:30,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:30,501][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442764 seconds
+[2025-12-18 15:25:30,776][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391378 seconds
+[2025-12-18 15:25:30,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427274 seconds
+[2025-12-18 15:25:30,854][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:30,902][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:31,027][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440411 seconds
+[2025-12-18 15:25:31,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:31,300][openai._base_client][INFO] - Retrying request to /chat/completions in 0.759743 seconds
+[2025-12-18 15:25:31,517][openai._base_client][INFO] - Retrying request to /chat/completions in 0.936188 seconds
+[2025-12-18 15:25:31,607][openai._base_client][INFO] - Retrying request to /chat/completions in 0.827594 seconds
+[2025-12-18 15:25:31,734][openai._base_client][INFO] - Retrying request to /chat/completions in 0.889141 seconds
+[2025-12-18 15:25:31,760][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:31,817][openai._base_client][INFO] - Retrying request to /chat/completions in 0.925572 seconds
+[2025-12-18 15:25:31,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:31,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:33,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:33,149][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:33,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:33,892][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:33,902][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:34,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:34,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:36,748][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:36,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:37,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:42,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457535 seconds
+[2025-12-18 15:25:42,616][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438451 seconds
+[2025-12-18 15:25:42,618][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457125 seconds
+[2025-12-18 15:25:42,619][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408414 seconds
+[2025-12-18 15:25:42,620][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421136 seconds
+[2025-12-18 15:25:42,620][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448641 seconds
+[2025-12-18 15:25:42,621][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424778 seconds
+[2025-12-18 15:25:42,621][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409978 seconds
+[2025-12-18 15:25:42,622][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484409 seconds
+[2025-12-18 15:25:42,623][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458714 seconds
+[2025-12-18 15:25:42,623][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477716 seconds
+[2025-12-18 15:25:42,625][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408268 seconds
+[2025-12-18 15:25:42,626][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470951 seconds
+[2025-12-18 15:25:42,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474656 seconds
+[2025-12-18 15:25:42,681][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469622 seconds
+[2025-12-18 15:25:43,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.983515 seconds
+[2025-12-18 15:25:43,384][openai._base_client][INFO] - Retrying request to /chat/completions in 0.795317 seconds
+[2025-12-18 15:25:43,385][openai._base_client][INFO] - Retrying request to /chat/completions in 0.995644 seconds
+[2025-12-18 15:25:43,388][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863953 seconds
+[2025-12-18 15:25:43,420][openai._base_client][INFO] - Retrying request to /chat/completions in 0.848103 seconds
+[2025-12-18 15:25:43,426][openai._base_client][INFO] - Retrying request to /chat/completions in 0.770467 seconds
+[2025-12-18 15:25:43,453][openai._base_client][INFO] - Retrying request to /chat/completions in 0.924473 seconds
+[2025-12-18 15:25:43,455][openai._base_client][INFO] - Retrying request to /chat/completions in 0.918198 seconds
+[2025-12-18 15:25:45,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408903 seconds
+[2025-12-18 15:25:45,085][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409995 seconds
+[2025-12-18 15:25:46,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:47,147][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:47,228][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:47,280][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:48,085][openai._base_client][INFO] - Retrying request to /chat/completions in 0.800452 seconds
+[2025-12-18 15:25:48,110][openai._base_client][INFO] - Retrying request to /chat/completions in 0.963124 seconds
+[2025-12-18 15:25:48,357][openai._base_client][INFO] - Retrying request to /chat/completions in 0.909467 seconds
+[2025-12-18 15:25:48,371][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:48,713][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:48,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:48,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:49,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:49,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:50,497][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816514 seconds
+[2025-12-18 15:25:50,851][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:50,934][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:50,988][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:51,568][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:51,796][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:51,920][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:52,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:52,312][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:52,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:52,525][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:53,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:54,049][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410261 seconds
+[2025-12-18 15:25:54,204][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436278 seconds
+[2025-12-18 15:25:54,241][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453216 seconds
+[2025-12-18 15:25:54,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443437 seconds
+[2025-12-18 15:25:54,340][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:54,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:54,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:54,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:55,123][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:55,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:55,207][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:55,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:55,495][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:55,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:56,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:57,716][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:57,719][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:57,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:57,925][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:57,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:57,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:58,107][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:58,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:58,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:58,363][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:58,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:58,754][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:59,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:25:59,275][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458027 seconds
+[2025-12-18 15:25:59,644][openai._base_client][INFO] - Retrying request to /chat/completions in 0.778837 seconds
+[2025-12-18 15:25:59,868][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:00,567][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:00,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:00,901][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:01,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:01,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:01,073][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:02,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:02,230][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:02,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:02,780][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:03,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:03,776][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:03,805][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:04,513][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,392][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:05,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:06,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:06,874][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:06,949][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:07,089][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:07,672][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:08,089][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:08,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:08,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:08,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:09,039][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:14,049][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475810 seconds
+[2025-12-18 15:26:17,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:17,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:17,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:17,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:17,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,432][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,681][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:18,818][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:19,531][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899809 seconds
+[2025-12-18 15:26:19,563][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:19,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:19,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:19,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:19,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:20,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,250][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,404][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,418][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:21,848][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:22,284][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:22,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:22,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:22,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:22,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:22,577][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:23,389][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,266][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,396][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:24,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:25,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:25,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:25,121][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:25,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:25,177][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:25,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:26,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,244][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,272][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:27,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:28,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:28,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:28,389][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:28,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:29,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:29,874][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:29,901][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:29,917][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:29,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:29,989][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:30,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:30,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:30,438][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403291 seconds
+[2025-12-18 15:26:30,479][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:30,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:30,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:31,058][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:31,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:31,170][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:31,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:32,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:32,275][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:32,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:32,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:32,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:32,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:33,666][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:38,676][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464815 seconds
+[2025-12-18 15:26:41,223][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:41,339][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:41,479][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:41,500][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:41,518][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:41,544][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:42,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:42,204][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:42,294][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:42,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:42,333][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:43,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426372 seconds
+[2025-12-18 15:26:43,699][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416494 seconds
+[2025-12-18 15:26:43,700][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429844 seconds
+[2025-12-18 15:26:43,700][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421441 seconds
+[2025-12-18 15:26:43,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:43,881][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:44,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:44,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:44,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:44,147][openai._base_client][INFO] - Retrying request to /chat/completions in 0.997360 seconds
+[2025-12-18 15:26:44,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:44,742][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:44,868][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:45,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:45,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:45,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:46,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:46,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:46,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:46,895][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:47,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:47,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:47,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:48,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:48,004][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:48,522][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:48,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:48,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:49,119][openai._base_client][INFO] - Retrying request to /chat/completions in 0.798287 seconds
+[2025-12-18 15:26:49,125][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916916 seconds
+[2025-12-18 15:26:49,127][openai._base_client][INFO] - Retrying request to /chat/completions in 0.846464 seconds
+[2025-12-18 15:26:49,131][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783954 seconds
+[2025-12-18 15:26:49,533][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:49,585][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:49,645][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:50,195][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:50,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:50,934][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:51,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:51,208][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:51,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:51,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:52,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:52,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:53,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:53,170][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:53,373][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:53,961][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:54,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:54,640][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:54,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:54,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:54,884][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:55,151][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469818 seconds
+[2025-12-18 15:26:55,685][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:55,757][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:56,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:56,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:56,611][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:56,651][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:57,588][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:58,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:58,163][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:58,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:58,317][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:59,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:59,350][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:59,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:59,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:59,909][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:59,923][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488498 seconds
+[2025-12-18 15:26:59,924][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431696 seconds
+[2025-12-18 15:26:59,971][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:26:59,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412779 seconds
+[2025-12-18 15:27:00,650][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:00,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:01,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:01,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:01,593][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:02,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:04,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:04,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:04,529][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:09,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411371 seconds
+[2025-12-18 15:27:12,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,472][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,727][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,757][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:12,779][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:13,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:13,771][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:14,558][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436717 seconds
+[2025-12-18 15:27:14,559][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417123 seconds
+[2025-12-18 15:27:14,559][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416896 seconds
+[2025-12-18 15:27:14,560][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427805 seconds
+[2025-12-18 15:27:14,561][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414000 seconds
+[2025-12-18 15:27:15,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,693][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,757][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,770][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,775][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:15,783][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:16,476][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:16,916][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,699][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,723][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,780][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:18,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:19,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:19,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:19,783][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:19,980][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803142 seconds
+[2025-12-18 15:27:19,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920704 seconds
+[2025-12-18 15:27:19,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:19,990][openai._base_client][INFO] - Retrying request to /chat/completions in 0.790467 seconds
+[2025-12-18 15:27:21,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:21,253][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:21,285][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:21,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:21,356][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:21,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:21,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:21,408][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:22,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:22,656][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:22,882][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:22,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:24,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:24,418][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:24,498][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:24,690][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:24,827][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:25,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:25,164][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:25,194][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:25,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:25,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:25,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:26,392][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:26,573][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:26,585][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:27,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:27,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:27,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,181][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,187][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,265][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,694][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,869][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:28,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:30,584][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:31,177][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465209 seconds
+[2025-12-18 15:27:31,351][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:31,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:31,833][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:31,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:31,885][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:31,953][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:32,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:34,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:49,948][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:54,960][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461277 seconds
+[2025-12-18 15:27:59,164][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,223][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,349][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,434][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:27:59,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408693 seconds
+[2025-12-18 15:27:59,982][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413879 seconds
+[2025-12-18 15:27:59,983][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449235 seconds
+[2025-12-18 15:27:59,983][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448335 seconds
+[2025-12-18 15:27:59,984][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480358 seconds
+[2025-12-18 15:27:59,985][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455418 seconds
+[2025-12-18 15:28:00,425][openai._base_client][INFO] - Retrying request to /chat/completions in 0.796517 seconds
+[2025-12-18 15:28:01,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:03,224][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:03,426][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:03,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:03,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:03,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:03,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:03,709][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:04,258][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:04,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:05,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:05,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:05,394][openai._base_client][INFO] - Retrying request to /chat/completions in 0.975315 seconds
+[2025-12-18 15:28:05,435][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904695 seconds
+[2025-12-18 15:28:05,467][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806286 seconds
+[2025-12-18 15:28:05,676][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:07,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:07,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:07,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:07,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:07,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:07,663][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:08,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:08,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:09,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:09,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:09,056][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:09,769][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:11,228][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:11,231][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395492 seconds
+[2025-12-18 15:28:11,323][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:11,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:11,375][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:11,976][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:11,984][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:12,184][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:12,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:13,115][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:13,256][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:13,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:13,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:15,394][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:16,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422598 seconds
+[2025-12-18 15:28:16,347][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434840 seconds
+[2025-12-18 15:28:16,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:16,356][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:16,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:16,493][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:16,558][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:16,630][openai._base_client][INFO] - Retrying request to /chat/completions in 0.886825 seconds
+[2025-12-18 15:28:17,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:17,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:17,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:18,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:18,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:18,931][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:19,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:21,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:21,707][openai._base_client][INFO] - Retrying request to /chat/completions in 0.890929 seconds
+[2025-12-18 15:28:21,754][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:21,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:22,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:22,589][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:22,838][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:22,881][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:23,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:23,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:23,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:24,149][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:24,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:24,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:25,491][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:25,567][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:28,508][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:28,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:28,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:29,130][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:29,666][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:30,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-28-52/.hydra/config.yaml b/outputs/2025-12-18/15-28-52/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..308d390bba29e1ff02e98f25d11dcdaf349575d7
--- /dev/null
+++ b/outputs/2025-12-18/15-28-52/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-28-52/.hydra/hydra.yaml b/outputs/2025-12-18/15-28-52/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f2cb05f62df5dfb33379bbcbefafbbb1a6e98dc9
--- /dev/null
+++ b/outputs/2025-12-18/15-28-52/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-28-52
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-28-52/.hydra/overrides.yaml b/outputs/2025-12-18/15-28-52/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-28-52/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-28-52/eval_api.log b/outputs/2025-12-18/15-28-52/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..c1a2050fe96e30b3fc74e9d9fd3a7dab1f976454
--- /dev/null
+++ b/outputs/2025-12-18/15-28-52/eval_api.log
@@ -0,0 +1,448 @@
+[2025-12-18 15:28:59,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:28:59,571][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386102 seconds
+[2025-12-18 15:28:59,572][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420183 seconds
+[2025-12-18 15:28:59,572][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440818 seconds
+[2025-12-18 15:28:59,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375828 seconds
+[2025-12-18 15:28:59,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389241 seconds
+[2025-12-18 15:28:59,574][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445698 seconds
+[2025-12-18 15:28:59,574][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472192 seconds
+[2025-12-18 15:28:59,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490282 seconds
+[2025-12-18 15:28:59,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450198 seconds
+[2025-12-18 15:28:59,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390250 seconds
+[2025-12-18 15:29:00,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:00,090][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:00,092][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:00,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:00,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:04,965][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933082 seconds
+[2025-12-18 15:29:04,965][openai._base_client][INFO] - Retrying request to /chat/completions in 0.917771 seconds
+[2025-12-18 15:29:04,968][openai._base_client][INFO] - Retrying request to /chat/completions in 0.830329 seconds
+[2025-12-18 15:29:04,996][openai._base_client][INFO] - Retrying request to /chat/completions in 0.792687 seconds
+[2025-12-18 15:29:05,017][openai._base_client][INFO] - Retrying request to /chat/completions in 0.912229 seconds
+[2025-12-18 15:29:05,028][openai._base_client][INFO] - Retrying request to /chat/completions in 0.968513 seconds
+[2025-12-18 15:29:05,051][openai._base_client][INFO] - Retrying request to /chat/completions in 0.962668 seconds
+[2025-12-18 15:29:05,067][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905425 seconds
+[2025-12-18 15:29:05,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:06,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:11,563][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:16,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410955 seconds
+[2025-12-18 15:29:21,601][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423784 seconds
+[2025-12-18 15:29:21,602][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395973 seconds
+[2025-12-18 15:29:21,602][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391483 seconds
+[2025-12-18 15:29:21,603][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419255 seconds
+[2025-12-18 15:29:21,603][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492070 seconds
+[2025-12-18 15:29:21,603][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419711 seconds
+[2025-12-18 15:29:21,604][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387342 seconds
+[2025-12-18 15:29:21,604][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410783 seconds
+[2025-12-18 15:29:21,605][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480877 seconds
+[2025-12-18 15:29:21,605][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443100 seconds
+[2025-12-18 15:29:21,605][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375304 seconds
+[2025-12-18 15:29:21,606][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429294 seconds
+[2025-12-18 15:29:21,606][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421920 seconds
+[2025-12-18 15:29:21,990][openai._base_client][INFO] - Retrying request to /chat/completions in 0.974424 seconds
+[2025-12-18 15:29:26,987][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999235 seconds
+[2025-12-18 15:29:26,994][openai._base_client][INFO] - Retrying request to /chat/completions in 0.998661 seconds
+[2025-12-18 15:29:26,996][openai._base_client][INFO] - Retrying request to /chat/completions in 0.959288 seconds
+[2025-12-18 15:29:27,000][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842639 seconds
+[2025-12-18 15:29:27,018][openai._base_client][INFO] - Retrying request to /chat/completions in 0.958520 seconds
+[2025-12-18 15:29:27,025][openai._base_client][INFO] - Retrying request to /chat/completions in 0.813702 seconds
+[2025-12-18 15:29:27,025][openai._base_client][INFO] - Retrying request to /chat/completions in 0.998997 seconds
+[2025-12-18 15:29:27,027][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871631 seconds
+[2025-12-18 15:29:27,038][openai._base_client][INFO] - Retrying request to /chat/completions in 0.858684 seconds
+[2025-12-18 15:29:27,051][openai._base_client][INFO] - Retrying request to /chat/completions in 0.771735 seconds
+[2025-12-18 15:29:27,098][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979387 seconds
+[2025-12-18 15:29:27,293][openai._base_client][INFO] - Retrying request to /chat/completions in 0.824213 seconds
+[2025-12-18 15:29:27,598][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:33,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:33,890][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:33,892][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:34,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:39,619][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401124 seconds
+[2025-12-18 15:29:44,637][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388642 seconds
+[2025-12-18 15:29:44,638][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444722 seconds
+[2025-12-18 15:29:44,639][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438982 seconds
+[2025-12-18 15:29:44,639][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454830 seconds
+[2025-12-18 15:29:44,639][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452574 seconds
+[2025-12-18 15:29:44,835][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429322 seconds
+[2025-12-18 15:29:44,836][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383907 seconds
+[2025-12-18 15:29:45,055][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:45,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:45,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:45,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:45,214][openai._base_client][INFO] - Retrying request to /chat/completions in 0.992166 seconds
+[2025-12-18 15:29:45,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:49,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:49,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:50,030][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920371 seconds
+[2025-12-18 15:29:50,085][openai._base_client][INFO] - Retrying request to /chat/completions in 0.830081 seconds
+[2025-12-18 15:29:50,292][openai._base_client][INFO] - Retrying request to /chat/completions in 0.811696 seconds
+[2025-12-18 15:29:50,378][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:51,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:52,194][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:56,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:57,251][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:29:57,301][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:02,310][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432551 seconds
+[2025-12-18 15:30:07,320][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436078 seconds
+[2025-12-18 15:30:07,321][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406894 seconds
+[2025-12-18 15:30:07,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485641 seconds
+[2025-12-18 15:30:07,512][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463310 seconds
+[2025-12-18 15:30:07,520][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436613 seconds
+[2025-12-18 15:30:07,523][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475875 seconds
+[2025-12-18 15:30:08,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:09,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:12,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:13,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:13,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:13,475][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:13,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:13,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:14,434][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:19,446][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400789 seconds
+[2025-12-18 15:30:23,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:23,827][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:23,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:24,046][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:24,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:24,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:24,262][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:24,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:29,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495434 seconds
+[2025-12-18 15:30:34,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:34,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:34,147][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:34,857][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:35,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:35,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:38,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:43,173][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377223 seconds
+[2025-12-18 15:30:48,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393931 seconds
+[2025-12-18 15:30:57,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:57,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:57,222][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:57,292][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:57,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:57,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:30:58,180][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:03,189][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407351 seconds
+[2025-12-18 15:31:06,444][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:07,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:07,181][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:08,411][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468982 seconds
+[2025-12-18 15:31:08,412][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499457 seconds
+[2025-12-18 15:31:08,419][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384905 seconds
+[2025-12-18 15:31:08,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:11,598][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:13,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:13,598][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:18,607][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443659 seconds
+[2025-12-18 15:31:21,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:21,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:21,914][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:23,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483176 seconds
+[2025-12-18 15:31:23,821][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402945 seconds
+[2025-12-18 15:31:23,869][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454424 seconds
+[2025-12-18 15:31:24,325][openai._base_client][INFO] - Retrying request to /chat/completions in 0.847633 seconds
+[2025-12-18 15:31:26,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:27,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:28,236][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:32,679][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:37,690][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485572 seconds
+[2025-12-18 15:31:42,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:42,584][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:42,897][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481549 seconds
+[2025-12-18 15:31:42,908][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497323 seconds
+[2025-12-18 15:31:42,909][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422207 seconds
+[2025-12-18 15:31:43,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:44,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:46,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:48,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:48,415][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:53,425][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460009 seconds
+[2025-12-18 15:31:58,180][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:58,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:31:58,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404378 seconds
+[2025-12-18 15:31:59,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.789098 seconds
+[2025-12-18 15:32:01,983][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:02,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:02,978][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:04,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:04,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:09,663][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408296 seconds
+[2025-12-18 15:32:14,676][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462098 seconds
+[2025-12-18 15:32:14,677][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409614 seconds
+[2025-12-18 15:32:14,678][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432150 seconds
+[2025-12-18 15:32:14,870][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415651 seconds
+[2025-12-18 15:32:14,882][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413984 seconds
+[2025-12-18 15:32:19,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:20,093][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864071 seconds
+[2025-12-18 15:32:20,291][openai._base_client][INFO] - Retrying request to /chat/completions in 0.935138 seconds
+[2025-12-18 15:32:20,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:20,827][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:20,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:21,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:25,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:28,299][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:33,308][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393495 seconds
+[2025-12-18 15:32:38,088][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:38,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:38,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:38,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:38,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:38,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:38,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:43,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375437 seconds
+[2025-12-18 15:32:48,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405846 seconds
+[2025-12-18 15:32:48,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422048 seconds
+[2025-12-18 15:32:48,841][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484827 seconds
+[2025-12-18 15:32:48,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:49,208][openai._base_client][INFO] - Retrying request to /chat/completions in 0.968956 seconds
+[2025-12-18 15:32:49,232][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:49,280][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:53,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:53,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:32:56,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:01,462][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384129 seconds
+[2025-12-18 15:33:06,620][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:06,814][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:09,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:09,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:09,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:09,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:09,544][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:14,558][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497003 seconds
+[2025-12-18 15:33:19,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382854 seconds
+[2025-12-18 15:33:19,574][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469480 seconds
+[2025-12-18 15:33:20,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.934198 seconds
+[2025-12-18 15:33:20,365][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:20,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:24,964][openai._base_client][INFO] - Retrying request to /chat/completions in 0.847033 seconds
+[2025-12-18 15:33:25,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:25,048][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905813 seconds
+[2025-12-18 15:33:25,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:26,734][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:32,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:32,606][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:37,616][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441106 seconds
+[2025-12-18 15:33:42,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468925 seconds
+[2025-12-18 15:33:42,630][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435069 seconds
+[2025-12-18 15:33:43,908][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:43,918][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:44,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:45,194][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:45,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:48,072][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774153 seconds
+[2025-12-18 15:33:49,261][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:33:50,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:33:55,583][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447954 seconds
+[2025-12-18 15:34:00,597][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455777 seconds
+[2025-12-18 15:34:00,599][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449057 seconds
+[2025-12-18 15:34:00,599][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412020 seconds
+[2025-12-18 15:34:00,599][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437305 seconds
+[2025-12-18 15:34:01,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901478 seconds
+[2025-12-18 15:34:01,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:01,694][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:06,021][openai._base_client][INFO] - Retrying request to /chat/completions in 0.978058 seconds
+[2025-12-18 15:34:06,040][openai._base_client][INFO] - Retrying request to /chat/completions in 0.954680 seconds
+[2025-12-18 15:34:06,051][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899328 seconds
+[2025-12-18 15:34:06,056][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985550 seconds
+[2025-12-18 15:34:09,310][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:12,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:13,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:18,062][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476930 seconds
+[2025-12-18 15:34:23,074][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419224 seconds
+[2025-12-18 15:34:23,075][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419212 seconds
+[2025-12-18 15:34:23,076][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483194 seconds
+[2025-12-18 15:34:23,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.778039 seconds
+[2025-12-18 15:34:24,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:24,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:24,310][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:28,502][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807488 seconds
+[2025-12-18 15:34:28,504][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939238 seconds
+[2025-12-18 15:34:29,577][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:34:30,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:02,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:07,121][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403763 seconds
+[2025-12-18 15:35:12,115][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:12,128][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426179 seconds
+[2025-12-18 15:35:12,130][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380306 seconds
+[2025-12-18 15:35:12,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:12,329][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:12,530][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908260 seconds
+[2025-12-18 15:35:13,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:16,714][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:17,465][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:18,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:23,190][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437940 seconds
+[2025-12-18 15:35:28,033][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:28,199][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391228 seconds
+[2025-12-18 15:35:31,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:32,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:33,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:34,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:34,338][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:34,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:39,601][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482586 seconds
+[2025-12-18 15:35:43,439][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:44,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:44,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:45,089][openai._base_client][INFO] - Retrying request to /chat/completions in 0.852466 seconds
+[2025-12-18 15:35:45,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:45,559][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:46,024][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:48,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:53,818][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389094 seconds
+[2025-12-18 15:35:56,244][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:56,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:57,136][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:57,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:35:58,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485239 seconds
+[2025-12-18 15:35:58,829][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388947 seconds
+[2025-12-18 15:35:58,829][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420570 seconds
+[2025-12-18 15:36:00,943][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:01,954][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:02,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:07,050][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397700 seconds
+[2025-12-18 15:36:10,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:10,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:10,191][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:10,208][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:10,272][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:10,365][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:12,455][openai._base_client][INFO] - Retrying request to /chat/completions in 0.824617 seconds
+[2025-12-18 15:36:16,008][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:21,016][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383375 seconds
+[2025-12-18 15:36:23,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:24,482][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:26,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468810 seconds
+[2025-12-18 15:36:26,027][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403700 seconds
+[2025-12-18 15:36:26,027][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450625 seconds
+[2025-12-18 15:36:26,027][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394631 seconds
+[2025-12-18 15:36:26,404][openai._base_client][INFO] - Retrying request to /chat/completions in 0.754897 seconds
+[2025-12-18 15:36:28,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:28,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:31,483][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818939 seconds
+[2025-12-18 15:36:31,499][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949257 seconds
+[2025-12-18 15:36:34,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:34,465][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:39,475][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416748 seconds
+[2025-12-18 15:36:42,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:42,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:44,485][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399789 seconds
+[2025-12-18 15:36:44,486][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447851 seconds
+[2025-12-18 15:36:44,486][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437789 seconds
+[2025-12-18 15:36:44,896][openai._base_client][INFO] - Retrying request to /chat/completions in 0.857606 seconds
+[2025-12-18 15:36:45,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:46,761][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:48,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:49,940][openai._base_client][INFO] - Retrying request to /chat/completions in 0.989811 seconds
+[2025-12-18 15:36:50,338][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:52,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:57,721][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485085 seconds
+[2025-12-18 15:36:59,900][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:36:59,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:00,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:01,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:01,654][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:02,728][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473699 seconds
+[2025-12-18 15:37:02,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408917 seconds
+[2025-12-18 15:37:04,932][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:05,078][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:10,087][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394362 seconds
+[2025-12-18 15:37:12,424][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:13,052][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:14,091][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:15,096][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400754 seconds
+[2025-12-18 15:37:15,097][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476863 seconds
+[2025-12-18 15:37:15,097][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438019 seconds
+[2025-12-18 15:37:15,486][openai._base_client][INFO] - Retrying request to /chat/completions in 0.880767 seconds
+[2025-12-18 15:37:17,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:17,488][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:18,165][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:26,382][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444635 seconds
+[2025-12-18 15:37:28,752][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:28,850][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:29,047][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:29,054][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:31,394][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438901 seconds
+[2025-12-18 15:37:31,395][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414382 seconds
+[2025-12-18 15:37:31,396][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494130 seconds
+[2025-12-18 15:37:33,924][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:34,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:34,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:39,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413556 seconds
+[2025-12-18 15:37:42,856][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:44,480][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489675 seconds
+[2025-12-18 15:37:44,481][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380877 seconds
+[2025-12-18 15:37:44,481][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420775 seconds
+[2025-12-18 15:37:44,482][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468398 seconds
+[2025-12-18 15:37:44,482][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450207 seconds
+[2025-12-18 15:37:44,482][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415367 seconds
+[2025-12-18 15:37:47,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:47,165][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:49,903][openai._base_client][INFO] - Retrying request to /chat/completions in 0.982008 seconds
+[2025-12-18 15:37:49,904][openai._base_client][INFO] - Retrying request to /chat/completions in 0.987895 seconds
+[2025-12-18 15:37:49,954][openai._base_client][INFO] - Retrying request to /chat/completions in 0.909260 seconds
+[2025-12-18 15:37:49,974][openai._base_client][INFO] - Retrying request to /chat/completions in 0.753534 seconds
+[2025-12-18 15:37:52,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:37:52,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:00,903][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382289 seconds
+[2025-12-18 15:38:03,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:05,913][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434076 seconds
+[2025-12-18 15:38:05,915][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454944 seconds
+[2025-12-18 15:38:05,915][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490117 seconds
+[2025-12-18 15:38:05,915][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460409 seconds
+[2025-12-18 15:38:05,916][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479534 seconds
+[2025-12-18 15:38:06,289][openai._base_client][INFO] - Retrying request to /chat/completions in 0.922670 seconds
+[2025-12-18 15:38:08,109][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:11,375][openai._base_client][INFO] - Retrying request to /chat/completions in 0.759192 seconds
+[2025-12-18 15:38:11,379][openai._base_client][INFO] - Retrying request to /chat/completions in 0.946818 seconds
+[2025-12-18 15:38:11,399][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939273 seconds
+[2025-12-18 15:38:11,407][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767827 seconds
+[2025-12-18 15:38:13,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:22,350][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415762 seconds
+[2025-12-18 15:38:24,558][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:24,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:27,360][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470340 seconds
+[2025-12-18 15:38:27,361][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375806 seconds
+[2025-12-18 15:38:27,362][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424451 seconds
+[2025-12-18 15:38:27,363][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401018 seconds
+[2025-12-18 15:38:27,770][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783417 seconds
+[2025-12-18 15:38:30,206][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:30,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:30,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:32,792][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818970 seconds
+[2025-12-18 15:38:35,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:40,501][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465791 seconds
+[2025-12-18 15:38:43,341][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:43,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:43,957][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:44,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:44,467][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:44,523][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:46,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:51,998][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387160 seconds
+[2025-12-18 15:38:54,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:57,011][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427336 seconds
+[2025-12-18 15:38:57,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448347 seconds
+[2025-12-18 15:38:57,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377806 seconds
+[2025-12-18 15:38:57,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482652 seconds
+[2025-12-18 15:38:57,014][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403041 seconds
+[2025-12-18 15:38:57,391][openai._base_client][INFO] - Retrying request to /chat/completions in 0.865751 seconds
+[2025-12-18 15:38:59,112][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:59,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:38:59,930][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:00,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:00,473][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:02,501][openai._base_client][INFO] - Retrying request to /chat/completions in 0.893157 seconds
+[2025-12-18 15:39:05,192][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:10,203][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418002 seconds
+[2025-12-18 15:39:12,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:12,759][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:13,016][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:13,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:13,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:15,214][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376771 seconds
+[2025-12-18 15:39:15,216][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423334 seconds
+[2025-12-18 15:39:17,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:17,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-39-40/.hydra/config.yaml b/outputs/2025-12-18/15-39-40/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b384e47b36ec47243ec121c35af3a9ab7943f813
--- /dev/null
+++ b/outputs/2025-12-18/15-39-40/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 10
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 10
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-39-40/.hydra/hydra.yaml b/outputs/2025-12-18/15-39-40/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..27be748390565b33dff751d0a8747287fcccc419
--- /dev/null
+++ b/outputs/2025-12-18/15-39-40/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-39-40
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-39-40/.hydra/overrides.yaml b/outputs/2025-12-18/15-39-40/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-39-40/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-39-40/eval_api.log b/outputs/2025-12-18/15-39-40/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..a65134db28cf12f8d6f072f65a96a9255ec89583
--- /dev/null
+++ b/outputs/2025-12-18/15-39-40/eval_api.log
@@ -0,0 +1,91 @@
+[2025-12-18 15:39:44,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:44,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:45,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:45,140][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:45,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:45,284][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:46,529][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:46,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:47,458][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438974 seconds
+[2025-12-18 15:39:47,460][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404642 seconds
+[2025-12-18 15:39:49,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:50,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:55,408][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451791 seconds
+[2025-12-18 15:39:57,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:58,238][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:39:58,287][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:00,422][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491399 seconds
+[2025-12-18 15:40:00,423][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403430 seconds
+[2025-12-18 15:40:00,424][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463530 seconds
+[2025-12-18 15:40:00,424][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468473 seconds
+[2025-12-18 15:40:00,424][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479937 seconds
+[2025-12-18 15:40:02,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:02,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:03,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:04,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:05,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:10,671][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492101 seconds
+[2025-12-18 15:40:14,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:14,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:14,195][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:14,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:15,681][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404073 seconds
+[2025-12-18 15:40:15,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474949 seconds
+[2025-12-18 15:40:17,095][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:20,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:21,356][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:26,365][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486749 seconds
+[2025-12-18 15:40:30,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:30,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:30,979][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:31,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:31,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:32,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:37,846][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397405 seconds
+[2025-12-18 15:40:41,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:41,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:42,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:42,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:43,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:43,324][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:48,333][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423967 seconds
+[2025-12-18 15:40:52,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:53,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:53,761][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781698 seconds
+[2025-12-18 15:40:53,769][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:56,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:40:57,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:02,693][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440069 seconds
+[2025-12-18 15:41:06,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:06,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:06,533][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:07,250][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:08,161][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:13,170][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493375 seconds
+[2025-12-18 15:41:16,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:16,760][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:16,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:18,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:18,668][openai._base_client][INFO] - Retrying request to /chat/completions in 0.973641 seconds
+[2025-12-18 15:41:23,651][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:28,661][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407548 seconds
+[2025-12-18 15:41:30,868][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:31,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:31,515][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:33,671][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390324 seconds
+[2025-12-18 15:41:34,073][openai._base_client][INFO] - Retrying request to /chat/completions in 0.880427 seconds
+[2025-12-18 15:41:38,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:39,931][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:44,940][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483703 seconds
+[2025-12-18 15:41:49,951][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492564 seconds
+[2025-12-18 15:41:49,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408093 seconds
+[2025-12-18 15:41:49,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494236 seconds
+[2025-12-18 15:41:49,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464326 seconds
+[2025-12-18 15:41:50,427][openai._base_client][INFO] - Retrying request to /chat/completions in 0.991741 seconds
+[2025-12-18 15:41:55,367][openai._base_client][INFO] - Retrying request to /chat/completions in 0.950932 seconds
+[2025-12-18 15:41:55,420][openai._base_client][INFO] - Retrying request to /chat/completions in 0.934218 seconds
+[2025-12-18 15:41:55,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.868703 seconds
+[2025-12-18 15:41:55,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.927820 seconds
+[2025-12-18 15:41:56,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:41:59,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:42:00,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-44-06/.hydra/config.yaml b/outputs/2025-12-18/15-44-06/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b173723816aeddf15c4fe7f4999e85f7a01d956f
--- /dev/null
+++ b/outputs/2025-12-18/15-44-06/.hydra/config.yaml
@@ -0,0 +1,918 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 10
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 10
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+      max_retries: 5
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-44-06/.hydra/hydra.yaml b/outputs/2025-12-18/15-44-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..47079ebefadb4427a829dd0430175188932205ab
--- /dev/null
+++ b/outputs/2025-12-18/15-44-06/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-44-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-44-06/.hydra/overrides.yaml b/outputs/2025-12-18/15-44-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-44-06/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-44-06/eval_api.log b/outputs/2025-12-18/15-44-06/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/15-45-53/.hydra/config.yaml b/outputs/2025-12-18/15-45-53/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..681d8156bf6cd2fb1376b3c0378052dd0e5ce4b6
--- /dev/null
+++ b/outputs/2025-12-18/15-45-53/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 3600
diff --git a/outputs/2025-12-18/15-45-53/.hydra/hydra.yaml b/outputs/2025-12-18/15-45-53/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..02c003639cb8e90022e1e6493dd06cb804d389d5
--- /dev/null
+++ b/outputs/2025-12-18/15-45-53/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-45-53
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-45-53/.hydra/overrides.yaml b/outputs/2025-12-18/15-45-53/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-45-53/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-45-53/eval_api.log b/outputs/2025-12-18/15-45-53/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..168423cca5454c829ccd462a360baf26ff091ae5
--- /dev/null
+++ b/outputs/2025-12-18/15-45-53/eval_api.log
@@ -0,0 +1,3 @@
+[2025-12-18 15:46:03,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:46:09,011][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404544 seconds
+[2025-12-18 15:46:14,424][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788373 seconds
diff --git a/outputs/2025-12-18/15-46-33/.hydra/config.yaml b/outputs/2025-12-18/15-46-33/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3c7f04e9bb7a3b916dde72969acfbefcaf7ebad8
--- /dev/null
+++ b/outputs/2025-12-18/15-46-33/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-46-33/.hydra/hydra.yaml b/outputs/2025-12-18/15-46-33/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..da62a5b431c4b665ee988d0d5d0a8185a0e8084d
--- /dev/null
+++ b/outputs/2025-12-18/15-46-33/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-46-33
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-46-33/.hydra/overrides.yaml b/outputs/2025-12-18/15-46-33/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-46-33/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-46-33/eval_api.log b/outputs/2025-12-18/15-46-33/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..3208ea923e0f3438f903d3269cf7ce022383e767
--- /dev/null
+++ b/outputs/2025-12-18/15-46-33/eval_api.log
@@ -0,0 +1,6 @@
+[2025-12-18 15:46:39,616][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445067 seconds
+[2025-12-18 15:46:45,070][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788620 seconds
+[2025-12-18 15:47:00,880][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407138 seconds
+[2025-12-18 15:47:06,295][openai._base_client][INFO] - Retrying request to /chat/completions in 0.754279 seconds
+[2025-12-18 15:47:22,069][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437248 seconds
+[2025-12-18 15:47:27,516][openai._base_client][INFO] - Retrying request to /chat/completions in 0.854364 seconds
diff --git a/outputs/2025-12-18/15-47-45/.hydra/config.yaml b/outputs/2025-12-18/15-47-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3c7f04e9bb7a3b916dde72969acfbefcaf7ebad8
--- /dev/null
+++ b/outputs/2025-12-18/15-47-45/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-47-45/.hydra/hydra.yaml b/outputs/2025-12-18/15-47-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4378ae04b2c4473189aad6f31400d3b67460e59e
--- /dev/null
+++ b/outputs/2025-12-18/15-47-45/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-47-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-47-45/.hydra/overrides.yaml b/outputs/2025-12-18/15-47-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-47-45/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-47-45/eval_api.log b/outputs/2025-12-18/15-47-45/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..6d8e43162600fc526347502b9c52b91f6db35606
--- /dev/null
+++ b/outputs/2025-12-18/15-47-45/eval_api.log
@@ -0,0 +1,6 @@
+[2025-12-18 15:47:55,798][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7ef8aa292650>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 8ec93cde-2d17-40b2-a5cb-d7cbdd56f6ef)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 15:47:55,800][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 15:48:06,878][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7ef8aa291600>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: d880582b-daa9-4f1c-b7cf-c2cab825ab45)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 15:48:06,879][huggingface_hub.utils._http][WARNING] - Retrying in 2s [Retry 2/5].
+[2025-12-18 15:48:18,894][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7ef8aa293280>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 5efd7dd1-823b-4c6b-9bde-849b25d63075)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 15:48:18,897][huggingface_hub.utils._http][WARNING] - Retrying in 4s [Retry 3/5].
diff --git a/outputs/2025-12-18/15-48-35/.hydra/config.yaml b/outputs/2025-12-18/15-48-35/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9563b9a833d42bca9160e759ae50008e3e995d5a
--- /dev/null
+++ b/outputs/2025-12-18/15-48-35/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 2048
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-48-35/.hydra/hydra.yaml b/outputs/2025-12-18/15-48-35/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8b45a882d18a1cd4545d34f3a53166e949c51cd4
--- /dev/null
+++ b/outputs/2025-12-18/15-48-35/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-48-35
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-48-35/.hydra/overrides.yaml b/outputs/2025-12-18/15-48-35/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-48-35/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-48-35/eval_api.log b/outputs/2025-12-18/15-48-35/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..084704723fc12a9149ccac268a4a801a850bb3ec
--- /dev/null
+++ b/outputs/2025-12-18/15-48-35/eval_api.log
@@ -0,0 +1,5 @@
+[2025-12-18 15:48:47,714][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:48:47,750][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459024 seconds
+[2025-12-18 15:48:50,534][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:48:50,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419288 seconds
+[2025-12-18 15:48:53,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-50-09/.hydra/config.yaml b/outputs/2025-12-18/15-50-09/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6d4198493481ce9dceaaf4f045d5290d675875bd
--- /dev/null
+++ b/outputs/2025-12-18/15-50-09/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 2048
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-50-09/.hydra/hydra.yaml b/outputs/2025-12-18/15-50-09/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bd716dcb1c99691fbb19e145a699df3a9154d43f
--- /dev/null
+++ b/outputs/2025-12-18/15-50-09/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-50-09
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-50-09/.hydra/overrides.yaml b/outputs/2025-12-18/15-50-09/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-50-09/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-50-09/eval_api.log b/outputs/2025-12-18/15-50-09/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..8f59e3417b2958277e78d05c07dd6e262052f859
--- /dev/null
+++ b/outputs/2025-12-18/15-50-09/eval_api.log
@@ -0,0 +1,163 @@
+[2025-12-18 15:50:17,528][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376760 seconds
+[2025-12-18 15:50:22,333][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460960 seconds
+[2025-12-18 15:50:22,334][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383020 seconds
+[2025-12-18 15:50:22,334][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455229 seconds
+[2025-12-18 15:50:22,334][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386628 seconds
+[2025-12-18 15:50:22,334][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419591 seconds
+[2025-12-18 15:50:22,335][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426037 seconds
+[2025-12-18 15:50:22,335][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452122 seconds
+[2025-12-18 15:50:22,335][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467657 seconds
+[2025-12-18 15:50:22,336][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390923 seconds
+[2025-12-18 15:50:22,336][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424499 seconds
+[2025-12-18 15:50:22,336][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395460 seconds
+[2025-12-18 15:50:22,912][openai._base_client][INFO] - Retrying request to /chat/completions in 0.854592 seconds
+[2025-12-18 15:50:25,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:25,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:26,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:27,720][openai._base_client][INFO] - Retrying request to /chat/completions in 0.753346 seconds
+[2025-12-18 15:50:27,723][openai._base_client][INFO] - Retrying request to /chat/completions in 0.813426 seconds
+[2025-12-18 15:50:27,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.780527 seconds
+[2025-12-18 15:50:27,734][openai._base_client][INFO] - Retrying request to /chat/completions in 0.925520 seconds
+[2025-12-18 15:50:27,763][openai._base_client][INFO] - Retrying request to /chat/completions in 0.913258 seconds
+[2025-12-18 15:50:27,791][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818988 seconds
+[2025-12-18 15:50:27,806][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869901 seconds
+[2025-12-18 15:50:27,967][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:29,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:30,736][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:31,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:31,521][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:32,020][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:33,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:33,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:33,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:33,413][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:34,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:35,349][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:36,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:36,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:37,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:37,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:38,311][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:38,396][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:39,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:39,467][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:39,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:39,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:40,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:40,850][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:42,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:42,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:43,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:43,199][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:43,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:43,619][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:44,446][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:44,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:45,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:45,905][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:46,055][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:46,260][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:46,561][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:47,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:47,428][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:47,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:47,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:48,290][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:48,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:49,146][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:50,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:51,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:51,488][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:51,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:51,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:51,902][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:52,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:53,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:53,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:53,977][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:55,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:55,775][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:55,796][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:56,243][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:56,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:57,070][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:57,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:57,248][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:57,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:57,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:57,888][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:58,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:59,106][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:59,371][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:50:59,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:00,679][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:00,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:00,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:01,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:01,549][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:02,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:02,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:02,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:03,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:03,186][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:03,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:05,190][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:05,640][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:06,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:06,757][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:06,833][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:07,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:08,091][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:08,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:08,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:09,037][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:09,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:09,693][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:10,022][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:10,388][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:10,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:11,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:11,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:11,658][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:11,840][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:11,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:12,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:13,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:14,264][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:14,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:14,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:15,737][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:16,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:16,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:16,890][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:19,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:24,606][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:25,694][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:26,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:27,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:27,724][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:28,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:29,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:31,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:33,253][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:33,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:33,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:36,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:41,725][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464574 seconds
+[2025-12-18 15:51:46,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:48,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:50,536][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:52,769][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:54,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:57,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:51:57,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:52:02,719][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454483 seconds
+[2025-12-18 15:52:16,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:52:19,731][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:52:24,740][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392376 seconds
+[2025-12-18 15:52:29,745][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443261 seconds
+[2025-12-18 15:52:44,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:52:45,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:52:51,004][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391184 seconds
+[2025-12-18 15:53:06,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:53:11,177][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472650 seconds
diff --git a/outputs/2025-12-18/15-53-48/.hydra/config.yaml b/outputs/2025-12-18/15-53-48/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b4dc8a5d7e17d36acc8c418dc742840878832ac8
--- /dev/null
+++ b/outputs/2025-12-18/15-53-48/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-53-48/.hydra/hydra.yaml b/outputs/2025-12-18/15-53-48/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..20f3503071e70f166ece47f56d05086907a9c106
--- /dev/null
+++ b/outputs/2025-12-18/15-53-48/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-53-48
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-53-48/.hydra/overrides.yaml b/outputs/2025-12-18/15-53-48/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-53-48/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-53-48/eval_api.log b/outputs/2025-12-18/15-53-48/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..cf0de21110b3ef19ee167359602ebdbd3c2f3df9
--- /dev/null
+++ b/outputs/2025-12-18/15-53-48/eval_api.log
@@ -0,0 +1,132 @@
+[2025-12-18 15:53:59,134][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f124aa87970>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: fd28f0ad-28f8-4bc7-a40b-b631e1b0d607)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 15:53:59,136][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 15:54:05,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:05,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:06,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:06,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:06,558][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:06,761][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:06,785][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473948 seconds
+[2025-12-18 15:54:06,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467295 seconds
+[2025-12-18 15:54:06,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:07,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:08,209][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:08,311][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:08,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:09,226][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:09,517][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:09,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:10,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:10,536][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:10,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:11,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:11,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:12,263][openai._base_client][INFO] - Retrying request to /chat/completions in 0.761587 seconds
+[2025-12-18 15:54:13,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:13,663][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:13,998][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:14,796][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:15,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:15,329][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:15,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:15,993][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:16,313][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:16,798][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:17,284][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:19,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:25,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:26,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429355 seconds
+[2025-12-18 15:54:27,493][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:27,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:27,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:27,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:28,206][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:28,311][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:28,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:28,699][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:29,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:29,471][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:29,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:29,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:29,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:30,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:30,858][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:31,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:31,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:31,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:31,914][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:32,001][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:32,042][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:32,260][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:32,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:32,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:32,583][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:33,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:34,779][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:40,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:40,255][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474144 seconds
+[2025-12-18 15:54:44,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:44,585][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:45,255][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425327 seconds
+[2025-12-18 15:54:45,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474608 seconds
+[2025-12-18 15:54:45,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446415 seconds
+[2025-12-18 15:54:45,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399812 seconds
+[2025-12-18 15:54:45,257][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436246 seconds
+[2025-12-18 15:54:45,257][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411693 seconds
+[2025-12-18 15:54:45,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405128 seconds
+[2025-12-18 15:54:45,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450741 seconds
+[2025-12-18 15:54:45,260][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386975 seconds
+[2025-12-18 15:54:45,260][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489160 seconds
+[2025-12-18 15:54:45,260][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490560 seconds
+[2025-12-18 15:54:45,261][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433846 seconds
+[2025-12-18 15:54:45,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:45,873][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:45,916][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:47,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:47,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:48,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:50,669][openai._base_client][INFO] - Retrying request to /chat/completions in 0.801495 seconds
+[2025-12-18 15:54:50,672][openai._base_client][INFO] - Retrying request to /chat/completions in 0.824857 seconds
+[2025-12-18 15:54:50,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870729 seconds
+[2025-12-18 15:54:50,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.951659 seconds
+[2025-12-18 15:54:50,705][openai._base_client][INFO] - Retrying request to /chat/completions in 0.854888 seconds
+[2025-12-18 15:54:50,712][openai._base_client][INFO] - Retrying request to /chat/completions in 0.846068 seconds
+[2025-12-18 15:54:50,734][openai._base_client][INFO] - Retrying request to /chat/completions in 0.912538 seconds
+[2025-12-18 15:54:50,752][openai._base_client][INFO] - Retrying request to /chat/completions in 0.923449 seconds
+[2025-12-18 15:54:50,754][openai._base_client][INFO] - Retrying request to /chat/completions in 0.877300 seconds
+[2025-12-18 15:54:50,813][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:52,723][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:53,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:55,055][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:55,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:57,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:54:58,410][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:03,421][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376955 seconds
+[2025-12-18 15:55:08,431][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414504 seconds
+[2025-12-18 15:55:08,432][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381434 seconds
+[2025-12-18 15:55:08,801][openai._base_client][INFO] - Retrying request to /chat/completions in 0.821565 seconds
+[2025-12-18 15:55:11,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:13,851][openai._base_client][INFO] - Retrying request to /chat/completions in 0.960994 seconds
+[2025-12-18 15:55:17,262][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:22,271][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480093 seconds
+[2025-12-18 15:55:25,103][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:25,227][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451419 seconds
+[2025-12-18 15:55:26,688][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:26,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:27,048][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:27,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:28,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:29,763][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:30,228][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384379 seconds
+[2025-12-18 15:55:30,229][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430549 seconds
+[2025-12-18 15:55:30,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413882 seconds
+[2025-12-18 15:55:30,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420046 seconds
+[2025-12-18 15:55:30,231][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388252 seconds
+[2025-12-18 15:55:30,231][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402759 seconds
+[2025-12-18 15:55:30,683][openai._base_client][INFO] - Retrying request to /chat/completions in 0.796216 seconds
+[2025-12-18 15:55:32,154][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:32,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:33,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:33,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:33,978][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:55:34,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/15-55-47/.hydra/config.yaml b/outputs/2025-12-18/15-55-47/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6dfc05bb971ba706a90f75fb8754bbd694845419
--- /dev/null
+++ b/outputs/2025-12-18/15-55-47/.hydra/config.yaml
@@ -0,0 +1,918 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 8192
+      max_retries: 5
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-55-47/.hydra/hydra.yaml b/outputs/2025-12-18/15-55-47/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bdf58178338c7a82d4d1cf4e4969331ccc50f1d0
--- /dev/null
+++ b/outputs/2025-12-18/15-55-47/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-55-47
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-55-47/.hydra/overrides.yaml b/outputs/2025-12-18/15-55-47/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-55-47/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-55-47/eval_api.log b/outputs/2025-12-18/15-55-47/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/15-57-28/.hydra/config.yaml b/outputs/2025-12-18/15-57-28/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b4dc8a5d7e17d36acc8c418dc742840878832ac8
--- /dev/null
+++ b/outputs/2025-12-18/15-57-28/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-57-28/.hydra/hydra.yaml b/outputs/2025-12-18/15-57-28/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c41825551498845c87420a16e2dea864a5e1e5ab
--- /dev/null
+++ b/outputs/2025-12-18/15-57-28/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-57-28
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-57-28/.hydra/overrides.yaml b/outputs/2025-12-18/15-57-28/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-57-28/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-57-28/eval_api.log b/outputs/2025-12-18/15-57-28/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..4f4c7255e5bdefb9138fbc6c8b58859af6f58c09
--- /dev/null
+++ b/outputs/2025-12-18/15-57-28/eval_api.log
@@ -0,0 +1,56 @@
+[2025-12-18 15:57:35,379][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490816 seconds
+[2025-12-18 15:57:35,384][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491062 seconds
+[2025-12-18 15:57:37,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:37,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:38,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:38,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:39,160][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:39,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:39,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404583 seconds
+[2025-12-18 15:57:39,490][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:39,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:39,511][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474458 seconds
+[2025-12-18 15:57:39,993][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:39,994][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405651 seconds
+[2025-12-18 15:57:40,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:40,428][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417694 seconds
+[2025-12-18 15:57:40,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:40,875][openai._base_client][INFO] - Retrying request to /chat/completions in 0.796997 seconds
+[2025-12-18 15:57:40,878][openai._base_client][INFO] - Retrying request to /chat/completions in 0.907343 seconds
+[2025-12-18 15:57:42,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:42,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:42,444][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:42,447][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385608 seconds
+[2025-12-18 15:57:42,640][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:42,641][openai._base_client][INFO] - Retrying request to /chat/completions in 0.890662 seconds
+[2025-12-18 15:57:43,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:43,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:44,993][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:44,993][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462389 seconds
+[2025-12-18 15:57:45,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:46,900][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:48,236][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:48,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:48,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:48,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:48,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:48,681][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:48,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405036 seconds
+[2025-12-18 15:57:48,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:49,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:49,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:51,793][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451402 seconds
+[2025-12-18 15:57:52,745][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:53,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:54,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 15:57:54,072][openai._base_client][INFO] - Retrying request to /chat/completions in 0.919190 seconds
+[2025-12-18 15:57:54,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:54,858][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:55,264][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:56,033][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:56,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:57:58,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:01,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:04,047][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:09,057][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468679 seconds
+[2025-12-18 15:58:14,534][openai._base_client][INFO] - Retrying request to /chat/completions in 0.884013 seconds
diff --git a/outputs/2025-12-18/15-58-34/.hydra/config.yaml b/outputs/2025-12-18/15-58-34/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..29fb39cc76e08a3baa1ef122a57fb76cb16bc8b1
--- /dev/null
+++ b/outputs/2025-12-18/15-58-34/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 1024
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-58-34/.hydra/hydra.yaml b/outputs/2025-12-18/15-58-34/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..59560d8e8fd7c5bd2a155292b080926ef8902823
--- /dev/null
+++ b/outputs/2025-12-18/15-58-34/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-58-34
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-58-34/.hydra/overrides.yaml b/outputs/2025-12-18/15-58-34/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-58-34/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-58-34/eval_api.log b/outputs/2025-12-18/15-58-34/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..cf8a3590c77a96691acd8ab24c32815621cd9e67
--- /dev/null
+++ b/outputs/2025-12-18/15-58-34/eval_api.log
@@ -0,0 +1,50 @@
+[2025-12-18 15:58:38,764][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385764 seconds
+[2025-12-18 15:58:38,765][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433384 seconds
+[2025-12-18 15:58:38,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424742 seconds
+[2025-12-18 15:58:40,389][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:41,006][openai._base_client][INFO] - Retrying request to /chat/completions in 0.765863 seconds
+[2025-12-18 15:58:41,007][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944327 seconds
+[2025-12-18 15:58:41,016][openai._base_client][INFO] - Retrying request to /chat/completions in 0.861878 seconds
+[2025-12-18 15:58:41,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:41,701][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382136 seconds
+[2025-12-18 15:58:41,702][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456428 seconds
+[2025-12-18 15:58:41,702][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465238 seconds
+[2025-12-18 15:58:41,704][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495620 seconds
+[2025-12-18 15:58:41,704][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433526 seconds
+[2025-12-18 15:58:42,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:42,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:43,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:44,940][openai._base_client][INFO] - Retrying request to /chat/completions in 0.851629 seconds
+[2025-12-18 15:58:44,940][openai._base_client][INFO] - Retrying request to /chat/completions in 0.793919 seconds
+[2025-12-18 15:58:44,941][openai._base_client][INFO] - Retrying request to /chat/completions in 0.997361 seconds
+[2025-12-18 15:58:44,941][openai._base_client][INFO] - Retrying request to /chat/completions in 0.910586 seconds
+[2025-12-18 15:58:46,180][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:46,879][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:46,912][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:47,087][openai._base_client][INFO] - Retrying request to /chat/completions in 0.828477 seconds
+[2025-12-18 15:58:47,150][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498582 seconds
+[2025-12-18 15:58:48,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:48,044][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:49,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441344 seconds
+[2025-12-18 15:58:49,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:51,084][openai._base_client][INFO] - Retrying request to /chat/completions in 0.963709 seconds
+[2025-12-18 15:58:51,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:51,968][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:52,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:52,177][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:53,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:53,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:53,780][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:54,251][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:56,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:56,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:57,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:57,658][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:57,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:59,518][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:58:59,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:00,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:05,971][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472587 seconds
+[2025-12-18 15:59:06,572][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375852 seconds
+[2025-12-18 15:59:06,572][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463980 seconds
+[2025-12-18 15:59:06,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498948 seconds
diff --git a/outputs/2025-12-18/15-59-32/.hydra/config.yaml b/outputs/2025-12-18/15-59-32/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7d120667fbc3ddf0dbbd8ef54e236be8cb1d97c3
--- /dev/null
+++ b/outputs/2025-12-18/15-59-32/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 4096
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/15-59-32/.hydra/hydra.yaml b/outputs/2025-12-18/15-59-32/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fb7536b01a743354c63ae6b2c91b569470724de4
--- /dev/null
+++ b/outputs/2025-12-18/15-59-32/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/15-59-32
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/15-59-32/.hydra/overrides.yaml b/outputs/2025-12-18/15-59-32/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/15-59-32/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/15-59-32/eval_api.log b/outputs/2025-12-18/15-59-32/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..525c185b7b662f4e06d1452c04222f286c004335
--- /dev/null
+++ b/outputs/2025-12-18/15-59-32/eval_api.log
@@ -0,0 +1,88 @@
+[2025-12-18 15:59:36,503][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460989 seconds
+[2025-12-18 15:59:37,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:39,005][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:39,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:39,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:39,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.825466 seconds
+[2025-12-18 15:59:40,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:42,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:42,631][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:43,530][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:47,055][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:47,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:47,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:47,450][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:47,457][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:48,029][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:48,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:51,078][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:51,180][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489026 seconds
+[2025-12-18 15:59:51,182][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477913 seconds
+[2025-12-18 15:59:51,184][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462428 seconds
+[2025-12-18 15:59:51,185][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467437 seconds
+[2025-12-18 15:59:51,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458039 seconds
+[2025-12-18 15:59:51,187][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415110 seconds
+[2025-12-18 15:59:51,189][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403092 seconds
+[2025-12-18 15:59:51,190][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414499 seconds
+[2025-12-18 15:59:54,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:54,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:54,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:55,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.880643 seconds
+[2025-12-18 15:59:55,027][openai._base_client][INFO] - Retrying request to /chat/completions in 0.846396 seconds
+[2025-12-18 15:59:55,036][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451858 seconds
+[2025-12-18 15:59:55,037][openai._base_client][INFO] - Retrying request to /chat/completions in 0.843428 seconds
+[2025-12-18 15:59:55,038][openai._base_client][INFO] - Retrying request to /chat/completions in 0.963793 seconds
+[2025-12-18 15:59:55,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:56,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:56,190][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484089 seconds
+[2025-12-18 15:59:56,191][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437681 seconds
+[2025-12-18 15:59:56,595][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767746 seconds
+[2025-12-18 15:59:57,095][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:57,133][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 15:59:59,054][openai._base_client][INFO] - Retrying request to /chat/completions in 0.819796 seconds
+[2025-12-18 16:00:00,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:01,386][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:02,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:02,606][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:04,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:04,257][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:04,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:09,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456497 seconds
+[2025-12-18 16:00:13,560][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878859 seconds
+[2025-12-18 16:00:26,224][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:26,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418768 seconds
+[2025-12-18 16:00:29,005][openai._base_client][INFO] - Retrying request to /chat/completions in 0.836345 seconds
+[2025-12-18 16:00:29,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:29,076][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:29,195][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:30,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:30,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:31,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450268 seconds
+[2025-12-18 16:00:31,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420906 seconds
+[2025-12-18 16:00:31,596][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404431 seconds
+[2025-12-18 16:00:33,375][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:33,519][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:36,001][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:36,072][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449852 seconds
+[2025-12-18 16:00:36,074][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415413 seconds
+[2025-12-18 16:00:36,076][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487124 seconds
+[2025-12-18 16:00:38,221][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382057 seconds
+[2025-12-18 16:00:38,225][openai._base_client][INFO] - Retrying request to /chat/completions in 0.980102 seconds
+[2025-12-18 16:00:38,231][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472957 seconds
+[2025-12-18 16:00:38,364][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:39,213][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437576 seconds
+[2025-12-18 16:00:40,868][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:40,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:41,207][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:41,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:41,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:41,764][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415282 seconds
+[2025-12-18 16:00:41,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436772 seconds
+[2025-12-18 16:00:41,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479392 seconds
+[2025-12-18 16:00:41,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472908 seconds
+[2025-12-18 16:00:42,184][openai._base_client][INFO] - Retrying request to /chat/completions in 0.980028 seconds
+[2025-12-18 16:00:43,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:44,940][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:45,355][openai._base_client][INFO] - Retrying request to /chat/completions in 0.865476 seconds
+[2025-12-18 16:00:47,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:00:47,151][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/16-01-08/.hydra/config.yaml b/outputs/2025-12-18/16-01-08/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..733eb3ed5a96a8d749d99e026a2fa7d9fd653e75
--- /dev/null
+++ b/outputs/2025-12-18/16-01-08/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 4096
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/16-01-08/.hydra/hydra.yaml b/outputs/2025-12-18/16-01-08/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..01af3a41421fe4d55a655526860a7318d65f0ae9
--- /dev/null
+++ b/outputs/2025-12-18/16-01-08/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-01-08
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-01-08/.hydra/overrides.yaml b/outputs/2025-12-18/16-01-08/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-01-08/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-01-08/eval_api.log b/outputs/2025-12-18/16-01-08/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..7f4ed97b3d7ce9a46175011ce55eea661209dbce
--- /dev/null
+++ b/outputs/2025-12-18/16-01-08/eval_api.log
@@ -0,0 +1,600 @@
+[2025-12-18 16:01:19,049][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f8a56286620>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: e124f042-8ec1-4a8d-b817-89d8d68c79c5)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 16:01:19,051][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 16:01:25,515][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402832 seconds
+[2025-12-18 16:01:26,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:26,597][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:26,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:26,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:27,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:27,112][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:27,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:28,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:28,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:29,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:29,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:29,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:29,892][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:29,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:30,388][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:30,392][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:30,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:31,257][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:31,874][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:32,518][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:32,611][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:32,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:32,855][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:33,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:33,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:33,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:33,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:34,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:34,643][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:34,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:35,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:35,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:35,786][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:36,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:36,690][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:37,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:37,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:37,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:37,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:37,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:38,809][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:38,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:39,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:39,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:39,690][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:39,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:39,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:40,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:40,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:41,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:42,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:42,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:42,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:43,268][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:43,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:43,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:43,608][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:44,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:45,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:46,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:46,287][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:46,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:46,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:46,536][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:46,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:47,265][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:47,475][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:47,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:48,063][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:48,394][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:48,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:49,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:49,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:50,339][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:50,932][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:51,217][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:51,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:51,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:52,073][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:52,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:52,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:52,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:53,063][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:53,272][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:53,860][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:53,873][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:54,437][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:55,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:55,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:55,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:55,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:55,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:55,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:56,311][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:56,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:56,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:57,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:57,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:57,779][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:57,814][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:58,977][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:59,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:59,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:01:59,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:00,208][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:00,215][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:00,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:00,732][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:00,817][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:01,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:02,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:02,679][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:02,883][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:03,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:03,489][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:03,689][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:03,916][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:04,329][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:05,005][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:05,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:05,349][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:06,261][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:06,644][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:06,831][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:07,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:08,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:10,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:11,316][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:11,952][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426502 seconds
+[2025-12-18 16:02:11,972][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483824 seconds
+[2025-12-18 16:02:11,973][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412212 seconds
+[2025-12-18 16:02:11,974][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496547 seconds
+[2025-12-18 16:02:11,976][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385544 seconds
+[2025-12-18 16:02:13,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:13,879][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:13,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:14,059][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498149 seconds
+[2025-12-18 16:02:14,069][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489493 seconds
+[2025-12-18 16:02:14,405][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:14,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:15,004][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:15,012][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:15,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:15,230][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:15,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:15,913][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:15,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:16,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:16,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:16,500][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:16,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:16,979][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437592 seconds
+[2025-12-18 16:02:17,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:17,189][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:17,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:17,388][openai._base_client][INFO] - Retrying request to /chat/completions in 0.966998 seconds
+[2025-12-18 16:02:17,461][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833314 seconds
+[2025-12-18 16:02:17,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:18,208][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:18,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:18,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:18,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:18,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:19,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:19,559][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:19,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:20,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:20,418][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:20,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:20,869][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:20,952][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:21,199][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:21,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:21,518][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:21,521][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:21,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:21,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:22,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:22,477][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:22,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:22,713][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:22,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:22,885][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:23,206][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:23,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:23,673][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:23,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:24,177][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:24,523][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:24,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:24,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:24,805][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:24,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:24,998][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:25,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:25,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:25,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:25,742][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:25,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:26,513][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:26,933][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:26,983][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:26,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:27,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:27,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:27,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:27,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:27,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:27,897][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,154][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,650][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,673][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,719][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:28,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:29,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:29,734][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:30,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:30,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:30,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:30,645][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:30,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,185][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,215][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,254][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,439][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:31,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:32,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:32,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:32,722][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:32,869][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:33,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:33,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:33,567][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:34,404][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:34,468][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:35,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:35,669][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:37,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:37,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:38,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:38,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:38,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:39,695][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:42,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:47,888][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424744 seconds
+[2025-12-18 16:02:52,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:02:52,895][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445397 seconds
+[2025-12-18 16:02:54,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:00,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:01,405][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396395 seconds
+[2025-12-18 16:03:03,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:03,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:03,985][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:03,998][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:04,165][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:04,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:04,657][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:05,116][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:05,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:05,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:05,339][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:05,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:05,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:06,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:06,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:06,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:06,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:06,806][openai._base_client][INFO] - Retrying request to /chat/completions in 0.983940 seconds
+[2025-12-18 16:03:06,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:07,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:07,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:07,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:07,917][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:08,856][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:09,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:09,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:09,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:09,620][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:09,828][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:10,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:10,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:10,768][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:10,884][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:11,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:11,702][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,002][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:12,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:13,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:13,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:14,095][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:14,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:14,214][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:14,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:14,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:14,558][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:15,016][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:15,681][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:16,544][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:16,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:16,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:17,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:17,817][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:18,736][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:18,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:19,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:19,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:19,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:20,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:20,415][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:20,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:20,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:20,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:20,857][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:21,561][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,303][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,403][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:22,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:23,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:24,107][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:25,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:25,325][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:25,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:26,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:28,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:03:29,608][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:49,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:50,344][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391732 seconds
+[2025-12-18 16:05:51,997][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:52,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:52,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:52,656][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:52,691][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:52,908][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:52,919][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:52,976][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:53,115][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:53,238][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:53,260][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:53,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:54,035][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:54,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:54,164][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:54,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:55,044][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:55,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:55,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:55,345][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435882 seconds
+[2025-12-18 16:05:55,351][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387403 seconds
+[2025-12-18 16:05:55,353][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486032 seconds
+[2025-12-18 16:05:55,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:56,310][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:56,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:56,557][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:56,687][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:57,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:57,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:57,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:57,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:57,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:57,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:57,553][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,011][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,363][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,663][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,748][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:58,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:59,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:59,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:05:59,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:00,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:00,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:00,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:00,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:01,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:01,356][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:01,658][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:02,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:02,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:02,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:02,444][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:02,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:03,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:03,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:03,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:03,897][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:03,967][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:04,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:04,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:04,335][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:04,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:05,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:05,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:05,925][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:06,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:11,656][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:13,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:16,007][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:16,436][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385689 seconds
+[2025-12-18 16:06:16,437][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426689 seconds
+[2025-12-18 16:06:16,439][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386167 seconds
+[2025-12-18 16:06:18,059][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,444][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,711][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:18,985][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:19,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:19,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:19,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:19,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:20,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:20,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:20,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:21,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:21,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:21,301][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:21,439][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464919 seconds
+[2025-12-18 16:06:21,440][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420056 seconds
+[2025-12-18 16:06:21,442][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467872 seconds
+[2025-12-18 16:06:21,442][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432555 seconds
+[2025-12-18 16:06:21,443][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385771 seconds
+[2025-12-18 16:06:21,495][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:21,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.802661 seconds
+[2025-12-18 16:06:22,789][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:23,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:23,333][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:23,498][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:24,132][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:24,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:24,359][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:24,450][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:24,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:24,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:24,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:25,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:25,592][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:25,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:25,743][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:25,760][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:25,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:26,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:26,351][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:26,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:26,878][openai._base_client][INFO] - Retrying request to /chat/completions in 0.977299 seconds
+[2025-12-18 16:06:27,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:27,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:30,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:30,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:32,202][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:34,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:44,312][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:44,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:45,018][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457607 seconds
+[2025-12-18 16:06:45,020][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479465 seconds
+[2025-12-18 16:06:46,471][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:47,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:47,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:47,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:47,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:48,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:48,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:48,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:48,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:49,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:49,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:49,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:49,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:49,485][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:49,816][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:49,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:50,070][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:50,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:50,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:50,976][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:51,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:51,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:51,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:51,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:51,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:51,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:52,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:52,916][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:53,418][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:54,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:06:57,724][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:03,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:16,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:17,131][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408043 seconds
+[2025-12-18 16:07:18,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:18,644][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:18,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:18,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:19,317][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:19,437][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:19,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:19,794][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:20,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:20,642][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:21,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:21,426][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:22,136][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470879 seconds
+[2025-12-18 16:07:22,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.838746 seconds
+[2025-12-18 16:07:23,467][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:24,365][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:24,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:27,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:28,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:33,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:34,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:38,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:38,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402725 seconds
+[2025-12-18 16:07:38,571][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450053 seconds
+[2025-12-18 16:07:38,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379361 seconds
+[2025-12-18 16:07:40,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:40,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:40,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:40,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:41,011][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:41,022][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:43,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479511 seconds
+[2025-12-18 16:07:43,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445070 seconds
+[2025-12-18 16:07:43,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385519 seconds
+[2025-12-18 16:07:43,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429573 seconds
+[2025-12-18 16:07:43,579][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390155 seconds
+[2025-12-18 16:07:43,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939130 seconds
+[2025-12-18 16:07:43,955][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969404 seconds
+[2025-12-18 16:07:45,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:46,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:47,042][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:51,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:51,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:53,195][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:58,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:07:59,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:08:02,044][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:08:02,181][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494386 seconds
+[2025-12-18 16:08:02,182][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464409 seconds
+[2025-12-18 16:08:02,184][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428084 seconds
+[2025-12-18 16:08:05,202][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:08:07,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469972 seconds
+[2025-12-18 16:08:07,187][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494986 seconds
+[2025-12-18 16:08:07,187][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376948 seconds
+[2025-12-18 16:08:07,187][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386267 seconds
+[2025-12-18 16:08:07,616][openai._base_client][INFO] - Retrying request to /chat/completions in 0.981311 seconds
+[2025-12-18 16:08:07,649][openai._base_client][INFO] - Retrying request to /chat/completions in 0.813448 seconds
+[2025-12-18 16:08:12,580][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774135 seconds
+[2025-12-18 16:08:12,660][openai._base_client][INFO] - Retrying request to /chat/completions in 0.940551 seconds
+[2025-12-18 16:08:12,684][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797819 seconds
+[2025-12-18 16:08:13,859][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:08:13,908][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:08:41,611][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:08:46,622][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498637 seconds
+[2025-12-18 16:08:51,639][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464420 seconds
+[2025-12-18 16:08:51,640][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413116 seconds
+[2025-12-18 16:08:51,641][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387751 seconds
+[2025-12-18 16:08:52,125][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979249 seconds
+[2025-12-18 16:08:57,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.970529 seconds
+[2025-12-18 16:08:57,878][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:08:58,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:04,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:09,615][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398290 seconds
+[2025-12-18 16:09:15,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:15,737][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468011 seconds
+[2025-12-18 16:09:18,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:18,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:18,620][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:20,740][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391753 seconds
+[2025-12-18 16:09:21,208][openai._base_client][INFO] - Retrying request to /chat/completions in 0.758950 seconds
+[2025-12-18 16:09:22,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:23,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:24,944][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:09:25,494][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/16-15-01/.hydra/config.yaml b/outputs/2025-12-18/16-15-01/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe850a783d293ccc31fce1f8ecb9321726d06ff8
--- /dev/null
+++ b/outputs/2025-12-18/16-15-01/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 4096
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/16-15-01/.hydra/hydra.yaml b/outputs/2025-12-18/16-15-01/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fb349a5b318aa60a413d2cfa64cbd7acf6c8a725
--- /dev/null
+++ b/outputs/2025-12-18/16-15-01/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-15-01
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-15-01/.hydra/overrides.yaml b/outputs/2025-12-18/16-15-01/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-15-01/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-15-01/eval_api.log b/outputs/2025-12-18/16-15-01/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..e495664f76cad186c639142ed7c431892233e76d
--- /dev/null
+++ b/outputs/2025-12-18/16-15-01/eval_api.log
@@ -0,0 +1,393 @@
+[2025-12-18 16:15:11,292][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440056 seconds
+[2025-12-18 16:15:11,293][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388676 seconds
+[2025-12-18 16:15:11,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:12,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:13,103][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:13,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:13,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:13,680][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:14,140][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:14,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:14,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:16,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:16,685][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939395 seconds
+[2025-12-18 16:15:17,928][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:18,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:18,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:19,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:23,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:23,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:24,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:25,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:25,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:26,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:27,633][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494767 seconds
+[2025-12-18 16:15:29,724][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:30,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:33,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:34,419][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:34,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:35,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:35,685][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:36,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:37,107][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:37,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:42,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:42,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:44,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:44,809][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:45,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:47,403][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:47,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:48,519][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:48,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:50,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:51,476][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:52,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:52,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:56,363][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:57,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:57,725][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:58,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:58,934][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:15:59,563][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:01,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:05,063][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:05,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:05,910][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:06,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:06,186][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:06,317][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:08,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:09,112][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:09,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:10,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:11,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:15,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:15,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:15,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:16,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:16,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:17,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:19,112][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:19,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:19,304][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:21,342][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:22,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:22,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:22,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:22,827][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:25,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:27,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:29,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:29,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:30,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:31,127][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:31,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:31,170][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:31,751][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:32,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:33,725][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:34,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:35,037][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:35,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:35,878][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:36,113][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:36,222][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:39,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:40,714][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:41,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:41,450][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:41,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:41,667][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:43,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:43,267][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:44,906][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:45,177][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:45,517][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:47,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:48,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:48,915][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:49,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:49,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:49,690][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:49,803][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:50,251][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:51,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:52,052][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:52,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:52,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:52,985][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:53,945][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:55,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:56,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:56,731][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:57,521][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:16:58,663][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:01,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:02,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:02,457][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:06,338][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:07,212][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:12,223][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491775 seconds
+[2025-12-18 16:17:12,517][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445295 seconds
+[2025-12-18 16:17:12,528][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445845 seconds
+[2025-12-18 16:17:12,530][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461139 seconds
+[2025-12-18 16:17:12,926][openai._base_client][INFO] - Retrying request to /chat/completions in 0.892885 seconds
+[2025-12-18 16:17:13,165][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807706 seconds
+[2025-12-18 16:17:13,174][openai._base_client][INFO] - Retrying request to /chat/completions in 0.859121 seconds
+[2025-12-18 16:17:13,406][openai._base_client][INFO] - Retrying request to /chat/completions in 0.873953 seconds
+[2025-12-18 16:17:14,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380970 seconds
+[2025-12-18 16:17:14,675][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421981 seconds
+[2025-12-18 16:17:15,294][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844303 seconds
+[2025-12-18 16:17:17,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457809 seconds
+[2025-12-18 16:17:17,247][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447023 seconds
+[2025-12-18 16:17:17,247][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457793 seconds
+[2025-12-18 16:17:17,247][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402453 seconds
+[2025-12-18 16:17:17,248][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392281 seconds
+[2025-12-18 16:17:17,842][openai._base_client][INFO] - Retrying request to /chat/completions in 0.958349 seconds
+[2025-12-18 16:17:17,863][openai._base_client][INFO] - Retrying request to /chat/completions in 0.786956 seconds
+[2025-12-18 16:17:17,891][openai._base_client][INFO] - Retrying request to /chat/completions in 0.819659 seconds
+[2025-12-18 16:17:17,916][openai._base_client][INFO] - Retrying request to /chat/completions in 0.810747 seconds
+[2025-12-18 16:17:18,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:18,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:19,052][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444821 seconds
+[2025-12-18 16:17:19,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453827 seconds
+[2025-12-18 16:17:19,134][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390915 seconds
+[2025-12-18 16:17:19,206][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489747 seconds
+[2025-12-18 16:17:19,695][openai._base_client][INFO] - Retrying request to /chat/completions in 0.795394 seconds
+[2025-12-18 16:17:19,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908312 seconds
+[2025-12-18 16:17:19,785][openai._base_client][INFO] - Retrying request to /chat/completions in 0.911082 seconds
+[2025-12-18 16:17:19,895][openai._base_client][INFO] - Retrying request to /chat/completions in 0.794815 seconds
+[2025-12-18 16:17:21,094][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473634 seconds
+[2025-12-18 16:17:21,780][openai._base_client][INFO] - Retrying request to /chat/completions in 0.769723 seconds
+[2025-12-18 16:17:22,966][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415794 seconds
+[2025-12-18 16:17:23,596][openai._base_client][INFO] - Retrying request to /chat/completions in 0.775600 seconds
+[2025-12-18 16:17:24,783][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404525 seconds
+[2025-12-18 16:17:25,390][openai._base_client][INFO] - Retrying request to /chat/completions in 0.867162 seconds
+[2025-12-18 16:17:31,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:31,172][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:31,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:31,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:32,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:32,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:32,687][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:36,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:36,997][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:39,202][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:39,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:39,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:42,246][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:43,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:43,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:47,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:50,563][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:51,596][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:51,902][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:51,906][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:52,494][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:53,254][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:17:58,263][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384869 seconds
+[2025-12-18 16:17:59,116][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385378 seconds
+[2025-12-18 16:18:03,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419323 seconds
+[2025-12-18 16:18:03,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493432 seconds
+[2025-12-18 16:18:03,282][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452153 seconds
+[2025-12-18 16:18:03,282][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401193 seconds
+[2025-12-18 16:18:03,653][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812783 seconds
+[2025-12-18 16:18:03,882][openai._base_client][INFO] - Retrying request to /chat/completions in 0.837003 seconds
+[2025-12-18 16:18:03,933][openai._base_client][INFO] - Retrying request to /chat/completions in 0.957464 seconds
+[2025-12-18 16:18:04,215][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807886 seconds
+[2025-12-18 16:18:04,505][openai._base_client][INFO] - Retrying request to /chat/completions in 0.786969 seconds
+[2025-12-18 16:18:05,433][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:08,704][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901312 seconds
+[2025-12-18 16:18:09,667][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377275 seconds
+[2025-12-18 16:18:10,095][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399725 seconds
+[2025-12-18 16:18:10,224][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478609 seconds
+[2025-12-18 16:18:11,625][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421627 seconds
+[2025-12-18 16:18:14,726][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412660 seconds
+[2025-12-18 16:18:15,048][openai._base_client][INFO] - Retrying request to /chat/completions in 0.775717 seconds
+[2025-12-18 16:18:15,498][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815070 seconds
+[2025-12-18 16:18:15,707][openai._base_client][INFO] - Retrying request to /chat/completions in 0.802657 seconds
+[2025-12-18 16:18:16,496][openai._base_client][INFO] - Retrying request to /chat/completions in 0.879578 seconds
+[2025-12-18 16:18:16,518][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:16,595][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:16,843][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:17,050][openai._base_client][INFO] - Retrying request to /chat/completions in 0.993339 seconds
+[2025-12-18 16:18:17,111][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:17,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:18,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:18,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:18,626][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:19,616][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462140 seconds
+[2025-12-18 16:18:19,658][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:20,278][openai._base_client][INFO] - Retrying request to /chat/completions in 0.828933 seconds
+[2025-12-18 16:18:25,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:25,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464019 seconds
+[2025-12-18 16:18:26,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:29,282][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:30,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:31,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:32,687][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:33,657][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:34,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:35,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:35,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:36,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:18:41,525][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400933 seconds
+[2025-12-18 16:18:41,743][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375530 seconds
+[2025-12-18 16:18:42,125][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471425 seconds
+[2025-12-18 16:18:42,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.945651 seconds
+[2025-12-18 16:18:42,513][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920191 seconds
+[2025-12-18 16:18:42,874][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476072 seconds
+[2025-12-18 16:18:43,484][openai._base_client][INFO] - Retrying request to /chat/completions in 0.978280 seconds
+[2025-12-18 16:18:44,956][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416051 seconds
+[2025-12-18 16:18:45,985][openai._base_client][INFO] - Retrying request to /chat/completions in 0.866812 seconds
+[2025-12-18 16:18:46,540][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497077 seconds
+[2025-12-18 16:18:46,541][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385258 seconds
+[2025-12-18 16:18:46,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388879 seconds
+[2025-12-18 16:18:46,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473097 seconds
+[2025-12-18 16:18:46,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499458 seconds
+[2025-12-18 16:18:46,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409620 seconds
+[2025-12-18 16:18:46,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408888 seconds
+[2025-12-18 16:18:46,545][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479174 seconds
+[2025-12-18 16:18:46,546][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390054 seconds
+[2025-12-18 16:18:46,546][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434447 seconds
+[2025-12-18 16:18:46,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421481 seconds
+[2025-12-18 16:18:47,130][openai._base_client][INFO] - Retrying request to /chat/completions in 0.990154 seconds
+[2025-12-18 16:18:47,163][openai._base_client][INFO] - Retrying request to /chat/completions in 0.843830 seconds
+[2025-12-18 16:18:47,176][openai._base_client][INFO] - Retrying request to /chat/completions in 0.817519 seconds
+[2025-12-18 16:18:47,178][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869101 seconds
+[2025-12-18 16:18:47,223][openai._base_client][INFO] - Retrying request to /chat/completions in 0.945551 seconds
+[2025-12-18 16:18:47,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.982351 seconds
+[2025-12-18 16:18:47,649][openai._base_client][INFO] - Retrying request to /chat/completions in 0.849472 seconds
+[2025-12-18 16:18:48,350][openai._base_client][INFO] - Retrying request to /chat/completions in 0.769676 seconds
+[2025-12-18 16:18:48,538][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424559 seconds
+[2025-12-18 16:18:48,570][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421804 seconds
+[2025-12-18 16:18:48,622][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417839 seconds
+[2025-12-18 16:18:49,160][openai._base_client][INFO] - Retrying request to /chat/completions in 0.941027 seconds
+[2025-12-18 16:18:49,246][openai._base_client][INFO] - Retrying request to /chat/completions in 0.771184 seconds
+[2025-12-18 16:18:49,517][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377937 seconds
+[2025-12-18 16:18:50,096][openai._base_client][INFO] - Retrying request to /chat/completions in 0.910837 seconds
+[2025-12-18 16:18:50,510][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432897 seconds
+[2025-12-18 16:18:51,415][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452439 seconds
+[2025-12-18 16:18:51,568][openai._base_client][INFO] - Retrying request to /chat/completions in 0.971678 seconds
+[2025-12-18 16:18:52,081][openai._base_client][INFO] - Retrying request to /chat/completions in 0.829539 seconds
+[2025-12-18 16:18:52,957][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385518 seconds
+[2025-12-18 16:18:53,318][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487908 seconds
+[2025-12-18 16:18:53,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.902473 seconds
+[2025-12-18 16:18:54,076][openai._base_client][INFO] - Retrying request to /chat/completions in 0.775868 seconds
+[2025-12-18 16:18:55,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:05,815][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:13,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:13,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:15,561][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:18,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:19,517][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:20,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:20,906][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:24,642][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:26,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:26,831][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:28,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:28,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:41,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:19:46,325][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492255 seconds
+[2025-12-18 16:19:46,535][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429342 seconds
+[2025-12-18 16:19:46,538][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409576 seconds
+[2025-12-18 16:19:46,539][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466511 seconds
+[2025-12-18 16:19:46,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474112 seconds
+[2025-12-18 16:19:46,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449900 seconds
+[2025-12-18 16:19:46,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427592 seconds
+[2025-12-18 16:19:46,545][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413712 seconds
+[2025-12-18 16:19:46,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442553 seconds
+[2025-12-18 16:19:46,548][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436378 seconds
+[2025-12-18 16:19:46,549][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446932 seconds
+[2025-12-18 16:19:46,550][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418800 seconds
+[2025-12-18 16:19:46,552][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382533 seconds
+[2025-12-18 16:19:47,147][openai._base_client][INFO] - Retrying request to /chat/completions in 0.950701 seconds
+[2025-12-18 16:19:47,165][openai._base_client][INFO] - Retrying request to /chat/completions in 0.910868 seconds
+[2025-12-18 16:19:47,177][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904238 seconds
+[2025-12-18 16:19:47,179][openai._base_client][INFO] - Retrying request to /chat/completions in 0.778835 seconds
+[2025-12-18 16:19:47,182][openai._base_client][INFO] - Retrying request to /chat/completions in 0.974676 seconds
+[2025-12-18 16:19:47,196][openai._base_client][INFO] - Retrying request to /chat/completions in 0.763937 seconds
+[2025-12-18 16:19:47,199][openai._base_client][INFO] - Retrying request to /chat/completions in 0.941700 seconds
+[2025-12-18 16:19:47,228][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806963 seconds
+[2025-12-18 16:19:48,481][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469843 seconds
+[2025-12-18 16:19:49,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755952 seconds
+[2025-12-18 16:19:50,336][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450331 seconds
+[2025-12-18 16:19:50,983][openai._base_client][INFO] - Retrying request to /chat/completions in 0.768881 seconds
+[2025-12-18 16:19:51,821][openai._base_client][INFO] - Retrying request to /chat/completions in 0.804715 seconds
+[2025-12-18 16:19:51,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.915046 seconds
+[2025-12-18 16:19:51,962][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999718 seconds
+[2025-12-18 16:19:51,993][openai._base_client][INFO] - Retrying request to /chat/completions in 0.772863 seconds
+[2025-12-18 16:19:52,010][openai._base_client][INFO] - Retrying request to /chat/completions in 0.891707 seconds
+[2025-12-18 16:19:52,146][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482436 seconds
+[2025-12-18 16:19:53,163][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493865 seconds
+[2025-12-18 16:19:53,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399331 seconds
+[2025-12-18 16:19:53,345][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448299 seconds
+[2025-12-18 16:19:53,357][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426997 seconds
+[2025-12-18 16:19:53,459][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447530 seconds
+[2025-12-18 16:19:53,833][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844084 seconds
+[2025-12-18 16:19:53,866][openai._base_client][INFO] - Retrying request to /chat/completions in 0.881560 seconds
+[2025-12-18 16:19:53,991][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905222 seconds
+[2025-12-18 16:19:54,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.887318 seconds
+[2025-12-18 16:19:57,632][openai._base_client][INFO] - Retrying request to /chat/completions in 0.866230 seconds
+[2025-12-18 16:19:57,967][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410991 seconds
+[2025-12-18 16:19:58,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.830315 seconds
+[2025-12-18 16:19:58,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.858784 seconds
+[2025-12-18 16:20:15,609][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:15,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:16,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:21,730][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401968 seconds
+[2025-12-18 16:20:21,942][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422960 seconds
+[2025-12-18 16:20:21,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442670 seconds
+[2025-12-18 16:20:22,039][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493313 seconds
+[2025-12-18 16:20:22,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944443 seconds
+[2025-12-18 16:20:22,598][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972982 seconds
+[2025-12-18 16:20:26,746][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486961 seconds
+[2025-12-18 16:20:26,747][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494299 seconds
+[2025-12-18 16:20:27,138][openai._base_client][INFO] - Retrying request to /chat/completions in 0.805770 seconds
+[2025-12-18 16:20:49,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:49,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:50,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:50,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:50,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:50,878][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:50,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:51,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:51,963][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:52,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:52,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:52,724][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:53,154][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:54,649][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:56,013][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:20:57,257][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:12,419][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:12,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:14,073][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:14,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:14,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:14,883][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:14,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:15,001][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:15,146][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:21:20,156][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403160 seconds
+[2025-12-18 16:21:20,388][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422224 seconds
+[2025-12-18 16:21:20,390][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395531 seconds
+[2025-12-18 16:21:20,391][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481094 seconds
+[2025-12-18 16:21:20,403][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443775 seconds
+[2025-12-18 16:21:20,458][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418182 seconds
+[2025-12-18 16:21:20,463][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466552 seconds
+[2025-12-18 16:21:25,184][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482352 seconds
+[2025-12-18 16:21:25,185][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484085 seconds
+[2025-12-18 16:21:25,185][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440193 seconds
+[2025-12-18 16:21:25,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490154 seconds
+[2025-12-18 16:21:25,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420644 seconds
+[2025-12-18 16:21:25,567][openai._base_client][INFO] - Retrying request to /chat/completions in 0.962246 seconds
+[2025-12-18 16:21:25,850][openai._base_client][INFO] - Retrying request to /chat/completions in 0.958142 seconds
+[2025-12-18 16:21:25,878][openai._base_client][INFO] - Retrying request to /chat/completions in 0.923685 seconds
+[2025-12-18 16:21:25,933][openai._base_client][INFO] - Retrying request to /chat/completions in 0.795965 seconds
+[2025-12-18 16:21:30,611][openai._base_client][INFO] - Retrying request to /chat/completions in 0.841727 seconds
+[2025-12-18 16:21:30,671][openai._base_client][INFO] - Retrying request to /chat/completions in 0.790299 seconds
+[2025-12-18 16:21:30,672][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916592 seconds
diff --git a/outputs/2025-12-18/16-21-54/.hydra/config.yaml b/outputs/2025-12-18/16-21-54/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cd80d5e99f878c79b2ef823e04211cd190a258d8
--- /dev/null
+++ b/outputs/2025-12-18/16-21-54/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/16-21-54/.hydra/hydra.yaml b/outputs/2025-12-18/16-21-54/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..eab331e58a0b8f83a643d3b4c4678d1213630ac2
--- /dev/null
+++ b/outputs/2025-12-18/16-21-54/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-21-54
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-21-54/.hydra/overrides.yaml b/outputs/2025-12-18/16-21-54/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-21-54/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-21-54/eval_api.log b/outputs/2025-12-18/16-21-54/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..7de3f4980101c604367e21fca3836d73acdb2c86
--- /dev/null
+++ b/outputs/2025-12-18/16-21-54/eval_api.log
@@ -0,0 +1,179 @@
+[2025-12-18 16:22:03,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483357 seconds
+[2025-12-18 16:22:03,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438048 seconds
+[2025-12-18 16:22:03,239][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383220 seconds
+[2025-12-18 16:22:03,239][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482327 seconds
+[2025-12-18 16:22:03,239][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412039 seconds
+[2025-12-18 16:22:03,239][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492879 seconds
+[2025-12-18 16:22:03,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412690 seconds
+[2025-12-18 16:22:03,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481385 seconds
+[2025-12-18 16:22:03,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484927 seconds
+[2025-12-18 16:22:03,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480948 seconds
+[2025-12-18 16:22:03,241][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419426 seconds
+[2025-12-18 16:22:04,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:07,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:08,627][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818740 seconds
+[2025-12-18 16:22:08,655][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774444 seconds
+[2025-12-18 16:22:08,655][openai._base_client][INFO] - Retrying request to /chat/completions in 0.874750 seconds
+[2025-12-18 16:22:08,663][openai._base_client][INFO] - Retrying request to /chat/completions in 0.875156 seconds
+[2025-12-18 16:22:08,679][openai._base_client][INFO] - Retrying request to /chat/completions in 0.896340 seconds
+[2025-12-18 16:22:08,726][openai._base_client][INFO] - Retrying request to /chat/completions in 0.751891 seconds
+[2025-12-18 16:22:08,726][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816570 seconds
+[2025-12-18 16:22:08,726][openai._base_client][INFO] - Retrying request to /chat/completions in 0.868089 seconds
+[2025-12-18 16:22:08,727][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901468 seconds
+[2025-12-18 16:22:08,731][openai._base_client][INFO] - Retrying request to /chat/completions in 0.900266 seconds
+[2025-12-18 16:22:08,735][openai._base_client][INFO] - Retrying request to /chat/completions in 0.914002 seconds
+[2025-12-18 16:22:09,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:14,749][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:16,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:19,456][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492925 seconds
+[2025-12-18 16:22:19,536][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393846 seconds
+[2025-12-18 16:22:19,545][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448910 seconds
+[2025-12-18 16:22:19,583][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458912 seconds
+[2025-12-18 16:22:19,600][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464791 seconds
+[2025-12-18 16:22:19,634][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490088 seconds
+[2025-12-18 16:22:19,637][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445087 seconds
+[2025-12-18 16:22:19,656][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419074 seconds
+[2025-12-18 16:22:19,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:20,439][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:23,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:24,934][openai._base_client][INFO] - Retrying request to /chat/completions in 0.836403 seconds
+[2025-12-18 16:22:24,997][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905004 seconds
+[2025-12-18 16:22:25,079][openai._base_client][INFO] - Retrying request to /chat/completions in 0.840464 seconds
+[2025-12-18 16:22:25,084][openai._base_client][INFO] - Retrying request to /chat/completions in 0.841418 seconds
+[2025-12-18 16:22:25,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:26,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:27,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:27,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:29,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:33,903][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:34,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:35,022][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:36,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:37,796][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:41,702][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:44,013][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:44,254][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:45,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:46,351][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:47,912][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:49,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:50,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:51,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:51,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:52,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:56,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:56,069][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:57,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:22:57,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:00,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:02,815][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:03,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:03,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:04,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:04,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:04,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:06,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:07,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:08,291][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:08,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:09,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:09,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:10,824][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:11,161][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:13,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:13,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:13,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:14,883][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:15,172][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:16,221][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:19,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:20,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:22,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:22,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:22,999][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:24,060][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:24,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:29,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:31,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:32,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:33,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:33,944][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:34,877][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:35,158][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:36,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:37,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:38,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:38,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:40,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:40,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:40,897][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:40,976][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:42,284][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:46,809][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:48,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:48,464][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:49,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:51,774][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:52,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:53,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:54,172][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:55,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:56,070][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:57,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:58,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:59,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:59,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:59,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:23:59,846][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:01,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:02,031][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:03,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:04,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:04,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:04,989][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:05,242][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:05,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:06,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:08,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:09,170][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:14,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:16,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:19,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:20,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:25,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:27,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:27,720][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:29,131][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:40,733][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:41,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:46,980][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491186 seconds
+[2025-12-18 16:24:51,997][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475990 seconds
+[2025-12-18 16:24:51,998][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465355 seconds
+[2025-12-18 16:24:51,999][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488873 seconds
+[2025-12-18 16:24:52,476][openai._base_client][INFO] - Retrying request to /chat/completions in 0.978242 seconds
+[2025-12-18 16:24:54,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:54,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:56,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:57,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:57,684][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:57,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:24:58,364][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:25:01,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:25:01,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:25:08,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:25:42,997][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:25:51,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:25:55,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:25:59,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:26:04,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:26:05,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:26:14,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:26:15,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:26:17,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:26:22,272][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432905 seconds
+[2025-12-18 16:26:27,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419896 seconds
+[2025-12-18 16:26:32,709][openai._base_client][INFO] - Retrying request to /chat/completions in 0.964482 seconds
+[2025-12-18 16:27:02,908][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:27:15,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:27:20,398][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394452 seconds
+[2025-12-18 16:27:25,800][openai._base_client][INFO] - Retrying request to /chat/completions in 0.957019 seconds
diff --git a/outputs/2025-12-18/16-28-19/.hydra/config.yaml b/outputs/2025-12-18/16-28-19/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3fda6706afc5e78949175533edc6f1bb6811534f
--- /dev/null
+++ b/outputs/2025-12-18/16-28-19/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 10
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/16-28-19/.hydra/hydra.yaml b/outputs/2025-12-18/16-28-19/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..764a8221babb082cd3739e2481d800aceec5c713
--- /dev/null
+++ b/outputs/2025-12-18/16-28-19/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-28-19
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-28-19/.hydra/overrides.yaml b/outputs/2025-12-18/16-28-19/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-28-19/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-46-27/.hydra/config.yaml b/outputs/2025-12-18/16-46-27/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bd7e47599bde92b753c5e6e79d56fcf0ec1bf5a6
--- /dev/null
+++ b/outputs/2025-12-18/16-46-27/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/16-46-27/.hydra/hydra.yaml b/outputs/2025-12-18/16-46-27/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e7d0960cd9c6dfbf6207a039fb5577b8d44499c5
--- /dev/null
+++ b/outputs/2025-12-18/16-46-27/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-46-27
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-46-27/.hydra/overrides.yaml b/outputs/2025-12-18/16-46-27/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-46-27/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-46-27/eval_api.log b/outputs/2025-12-18/16-46-27/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..054ee534070e87042a1cdb96333b2e955d5fb429
--- /dev/null
+++ b/outputs/2025-12-18/16-46-27/eval_api.log
@@ -0,0 +1,5 @@
+[2025-12-18 16:46:34,781][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462039 seconds
+[2025-12-18 16:46:34,783][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455211 seconds
+[2025-12-18 16:46:34,783][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498694 seconds
+[2025-12-18 16:46:34,783][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495113 seconds
+[2025-12-18 16:46:34,784][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406860 seconds
diff --git a/outputs/2025-12-18/16-47-06/.hydra/hydra.yaml b/outputs/2025-12-18/16-47-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5e87536cfbfb7745168d2d9a09a4ce272f5dedb9
--- /dev/null
+++ b/outputs/2025-12-18/16-47-06/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-47-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-47-06/.hydra/overrides.yaml b/outputs/2025-12-18/16-47-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-47-06/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-47-06/eval_api.log b/outputs/2025-12-18/16-47-06/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..46dca9724e6b3e4291e82d14b55ac6abc646195b
--- /dev/null
+++ b/outputs/2025-12-18/16-47-06/eval_api.log
@@ -0,0 +1,75 @@
+[2025-12-18 16:47:16,499][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f8421787b80>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: d673bc47-c19a-4105-8467-a61d092d9bda)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 16:47:16,501][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 16:47:27,515][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f8421787a90>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: b00b349a-2602-418b-807d-fe0227cbc568)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 16:47:27,517][huggingface_hub.utils._http][WARNING] - Retrying in 2s [Retry 2/5].
+[2025-12-18 16:47:31,989][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458704 seconds
+[2025-12-18 16:47:36,762][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498189 seconds
+[2025-12-18 16:47:36,763][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475924 seconds
+[2025-12-18 16:47:36,764][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429260 seconds
+[2025-12-18 16:47:36,764][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473249 seconds
+[2025-12-18 16:47:36,764][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422991 seconds
+[2025-12-18 16:47:36,765][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384337 seconds
+[2025-12-18 16:47:36,765][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398725 seconds
+[2025-12-18 16:47:36,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433767 seconds
+[2025-12-18 16:47:36,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407852 seconds
+[2025-12-18 16:47:36,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443465 seconds
+[2025-12-18 16:47:36,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444148 seconds
+[2025-12-18 16:47:37,452][openai._base_client][INFO] - Retrying request to /chat/completions in 0.791156 seconds
+[2025-12-18 16:47:38,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:38,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:39,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:39,419][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:42,154][openai._base_client][INFO] - Retrying request to /chat/completions in 0.898974 seconds
+[2025-12-18 16:47:42,168][openai._base_client][INFO] - Retrying request to /chat/completions in 0.966377 seconds
+[2025-12-18 16:47:42,177][openai._base_client][INFO] - Retrying request to /chat/completions in 0.895886 seconds
+[2025-12-18 16:47:42,190][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869612 seconds
+[2025-12-18 16:47:42,196][openai._base_client][INFO] - Retrying request to /chat/completions in 0.836322 seconds
+[2025-12-18 16:47:42,202][openai._base_client][INFO] - Retrying request to /chat/completions in 0.927906 seconds
+[2025-12-18 16:47:42,214][openai._base_client][INFO] - Retrying request to /chat/completions in 0.923582 seconds
+[2025-12-18 16:47:42,215][openai._base_client][INFO] - Retrying request to /chat/completions in 0.891487 seconds
+[2025-12-18 16:47:42,241][openai._base_client][INFO] - Retrying request to /chat/completions in 0.787377 seconds
+[2025-12-18 16:47:42,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.769541 seconds
+[2025-12-18 16:47:42,264][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979561 seconds
+[2025-12-18 16:47:45,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:45,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:45,624][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:45,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:48,261][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486286 seconds
+[2025-12-18 16:47:53,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490509 seconds
+[2025-12-18 16:47:53,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431238 seconds
+[2025-12-18 16:47:53,038][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395475 seconds
+[2025-12-18 16:47:53,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475725 seconds
+[2025-12-18 16:47:53,066][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423677 seconds
+[2025-12-18 16:47:53,078][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399280 seconds
+[2025-12-18 16:47:53,112][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455705 seconds
+[2025-12-18 16:47:53,140][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431901 seconds
+[2025-12-18 16:47:53,143][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468088 seconds
+[2025-12-18 16:47:53,751][openai._base_client][INFO] - Retrying request to /chat/completions in 0.915320 seconds
+[2025-12-18 16:47:54,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:54,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:54,434][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:54,924][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:57,429][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:58,357][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:47:58,438][openai._base_client][INFO] - Retrying request to /chat/completions in 0.988540 seconds
+[2025-12-18 16:47:58,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.759653 seconds
+[2025-12-18 16:47:58,481][openai._base_client][INFO] - Retrying request to /chat/completions in 0.917648 seconds
+[2025-12-18 16:47:58,492][openai._base_client][INFO] - Retrying request to /chat/completions in 0.936763 seconds
+[2025-12-18 16:47:58,539][openai._base_client][INFO] - Retrying request to /chat/completions in 0.751448 seconds
+[2025-12-18 16:47:58,570][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969777 seconds
+[2025-12-18 16:47:58,576][openai._base_client][INFO] - Retrying request to /chat/completions in 0.765094 seconds
+[2025-12-18 16:47:58,615][openai._base_client][INFO] - Retrying request to /chat/completions in 0.948648 seconds
+[2025-12-18 16:48:00,771][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:01,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:01,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:01,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:01,905][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:04,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:04,673][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400740 seconds
+[2025-12-18 16:48:05,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:08,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:08,209][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:08,512][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:08,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:09,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:09,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/16-48-33/.hydra/config.yaml b/outputs/2025-12-18/16-48-33/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e394472378ff4c4b573d1c5d5a1ec267da0cc847
--- /dev/null
+++ b/outputs/2025-12-18/16-48-33/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 10
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/16-48-33/.hydra/hydra.yaml b/outputs/2025-12-18/16-48-33/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fb218a6e61ef745a0600de54ff7a8dbdfc3d22e0
--- /dev/null
+++ b/outputs/2025-12-18/16-48-33/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-48-33
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-48-33/.hydra/overrides.yaml b/outputs/2025-12-18/16-48-33/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-48-33/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-48-33/eval_api.log b/outputs/2025-12-18/16-48-33/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..ac4d20ce3090e389ff40d2df2a6c8a5d63f84da4
--- /dev/null
+++ b/outputs/2025-12-18/16-48-33/eval_api.log
@@ -0,0 +1,281 @@
+[2025-12-18 16:48:43,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:44,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:46,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:48,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:50,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:51,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:51,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:55,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:56,293][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:58,762][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:48:58,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:00,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:00,886][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:04,175][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:04,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:05,963][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:07,624][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:08,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:09,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:10,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:10,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:12,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:13,977][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:14,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:18,197][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:18,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:20,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:20,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:26,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:27,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:29,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:29,583][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:29,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:30,199][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:32,042][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:35,078][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:36,684][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:38,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:39,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:39,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:41,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:41,192][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:43,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:43,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:46,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:49,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:51,410][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:51,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:52,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:53,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:53,327][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:54,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:57,261][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:49:58,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:06,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:06,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:08,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:09,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:09,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:11,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:13,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:13,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:13,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:16,452][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:17,418][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:18,770][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:21,482][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:22,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:22,881][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:23,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:24,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:24,897][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:26,070][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:27,588][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:29,525][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:31,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:32,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:33,311][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:36,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:37,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:39,520][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:39,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:40,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:41,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:43,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:43,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:44,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:47,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:48,290][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:49,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:49,299][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:51,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:52,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:52,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:53,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:55,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:57,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:59,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:59,789][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:50:59,828][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:00,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:00,624][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:00,854][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:01,013][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:02,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:03,597][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:05,858][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:05,874][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:08,895][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:09,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:11,781][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:12,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:15,677][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:15,685][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:15,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:15,978][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:16,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:16,910][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:20,360][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:20,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:25,236][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:25,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:27,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:28,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:32,932][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:40,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:42,667][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:53,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:51:58,707][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380625 seconds
+[2025-12-18 16:52:11,372][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:11,653][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483934 seconds
+[2025-12-18 16:52:14,039][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:14,434][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:16,661][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428533 seconds
+[2025-12-18 16:52:16,662][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496138 seconds
+[2025-12-18 16:52:16,663][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427076 seconds
+[2025-12-18 16:52:16,663][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410874 seconds
+[2025-12-18 16:52:17,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:20,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:21,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:22,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:22,078][openai._base_client][INFO] - Retrying request to /chat/completions in 0.907736 seconds
+[2025-12-18 16:52:22,093][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949705 seconds
+[2025-12-18 16:52:22,094][openai._base_client][INFO] - Retrying request to /chat/completions in 0.990871 seconds
+[2025-12-18 16:52:22,413][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:25,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:28,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:32,999][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468585 seconds
+[2025-12-18 16:52:33,094][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454915 seconds
+[2025-12-18 16:52:37,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:37,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:38,475][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921928 seconds
+[2025-12-18 16:52:39,639][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:41,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:42,659][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:43,199][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:45,702][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:45,948][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:46,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:48,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:52,735][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:53,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:56,883][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:52:58,185][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:01,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:02,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:08,934][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:10,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:12,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:16,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:18,072][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:20,241][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:21,842][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:23,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:28,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:29,194][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:31,048][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:31,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:32,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:35,549][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:40,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:41,271][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:41,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:44,323][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:46,405][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:48,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:49,770][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:52,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:56,139][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:53:58,754][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:00,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:01,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:02,868][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:03,755][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:08,695][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:11,394][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:11,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:14,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:14,588][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:16,751][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:17,688][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:20,111][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:24,929][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:27,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:27,759][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:27,999][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:29,942][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:30,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:32,137][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:34,252][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:36,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:37,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:38,386][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:41,796][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:52,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:52,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:53,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:53,491][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:54,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:55,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:54:58,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:05,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:06,428][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:13,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:17,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:19,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:19,824][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:26,429][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:27,743][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:28,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:28,448][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:29,029][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:32,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:34,664][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:38,641][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:39,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:39,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:41,818][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:43,639][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:45,290][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:55:58,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:04,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:05,004][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:05,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:10,320][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463989 seconds
+[2025-12-18 16:56:15,350][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445539 seconds
+[2025-12-18 16:56:15,351][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492772 seconds
+[2025-12-18 16:56:15,352][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475219 seconds
+[2025-12-18 16:56:15,353][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467036 seconds
+[2025-12-18 16:56:15,354][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385550 seconds
+[2025-12-18 16:56:15,355][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463930 seconds
+[2025-12-18 16:56:15,355][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414857 seconds
+[2025-12-18 16:56:16,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:19,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:20,801][openai._base_client][INFO] - Retrying request to /chat/completions in 0.802944 seconds
+[2025-12-18 16:56:20,822][openai._base_client][INFO] - Retrying request to /chat/completions in 0.819521 seconds
+[2025-12-18 16:56:20,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.820955 seconds
+[2025-12-18 16:56:22,085][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:25,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:25,942][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:30,226][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:32,108][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:37,184][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:39,475][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:40,731][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:46,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:56:48,631][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:08,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:13,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:13,770][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:14,479][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:16,881][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:18,058][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:23,138][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450353 seconds
+[2025-12-18 16:57:28,152][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440630 seconds
+[2025-12-18 16:57:28,153][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422577 seconds
+[2025-12-18 16:57:28,153][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429691 seconds
+[2025-12-18 16:57:30,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:36,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:38,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:57:43,261][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/16-58-43/.hydra/config.yaml b/outputs/2025-12-18/16-58-43/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a678c0dbcc52c29fbbb53b54a08bb279d975b0da
--- /dev/null
+++ b/outputs/2025-12-18/16-58-43/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/16-58-43/.hydra/hydra.yaml b/outputs/2025-12-18/16-58-43/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..23b412d8df58b6d5487d1988c018c53abf2c6c59
--- /dev/null
+++ b/outputs/2025-12-18/16-58-43/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/16-58-43
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/16-58-43/.hydra/overrides.yaml b/outputs/2025-12-18/16-58-43/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/16-58-43/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/16-58-43/eval_api.log b/outputs/2025-12-18/16-58-43/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..2d115e8fc868c14544d8fb5d641a31348b8f3351
--- /dev/null
+++ b/outputs/2025-12-18/16-58-43/eval_api.log
@@ -0,0 +1,494 @@
+[2025-12-18 16:58:54,259][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f385ec6bc10>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 872e3c5b-4295-461b-a1d7-5d175ed086a8)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 16:58:54,260][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 16:59:02,442][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436272 seconds
+[2025-12-18 16:59:02,443][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391521 seconds
+[2025-12-18 16:59:02,446][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457342 seconds
+[2025-12-18 16:59:07,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949154 seconds
+[2025-12-18 16:59:07,882][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985867 seconds
+[2025-12-18 16:59:07,907][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908463 seconds
+[2025-12-18 16:59:08,066][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:08,111][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:08,931][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:09,457][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:09,531][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481931 seconds
+[2025-12-18 16:59:10,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:11,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:14,188][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394360 seconds
+[2025-12-18 16:59:14,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396184 seconds
+[2025-12-18 16:59:15,016][openai._base_client][INFO] - Retrying request to /chat/completions in 0.884465 seconds
+[2025-12-18 16:59:15,018][openai._base_client][INFO] - Retrying request to /chat/completions in 0.993074 seconds
+[2025-12-18 16:59:15,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:15,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:16,750][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414278 seconds
+[2025-12-18 16:59:18,763][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:19,587][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955056 seconds
+[2025-12-18 16:59:22,171][openai._base_client][INFO] - Retrying request to /chat/completions in 0.918318 seconds
+[2025-12-18 16:59:23,829][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498607 seconds
+[2025-12-18 16:59:25,908][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442022 seconds
+[2025-12-18 16:59:25,919][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492939 seconds
+[2025-12-18 16:59:26,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:27,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:27,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:28,902][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:29,331][openai._base_client][INFO] - Retrying request to /chat/completions in 0.897421 seconds
+[2025-12-18 16:59:29,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:31,355][openai._base_client][INFO] - Retrying request to /chat/completions in 0.966650 seconds
+[2025-12-18 16:59:36,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:36,942][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:37,158][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:38,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:39,111][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:40,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406708 seconds
+[2025-12-18 16:59:40,465][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:43,549][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:51,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:52,356][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:53,052][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:53,814][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:54,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:56,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 16:59:58,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:00,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:01,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:02,766][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:03,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:05,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:06,111][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:07,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:07,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:09,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:09,611][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:10,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:13,004][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:19,223][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:23,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:26,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:33,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:33,048][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:35,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:37,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:38,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:41,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:46,953][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:49,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:55,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:00:56,565][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:02,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:03,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:04,786][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 17:01:04,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412529 seconds
+[2025-12-18 17:01:06,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:07,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:08,416][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:10,335][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:11,219][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:13,963][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:15,745][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:16,879][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:20,609][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:20,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:23,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:23,921][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:25,285][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:27,186][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:27,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:29,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:30,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:30,568][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:32,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:36,642][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:37,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:39,092][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:44,224][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:44,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:46,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:46,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:46,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:47,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:47,733][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:50,657][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:51,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:55,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:55,704][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:55,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:58,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:01:59,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:00,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:00,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:00,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:00,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:03,887][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:04,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:05,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:05,982][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:07,389][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:09,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:12,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:14,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:15,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:16,589][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:18,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:20,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:23,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:29,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:29,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:31,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:32,672][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:34,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:34,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:36,394][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:37,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:38,180][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:45,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:51,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:54,731][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:02:59,795][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490426 seconds
+[2025-12-18 17:03:00,174][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442940 seconds
+[2025-12-18 17:03:00,227][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494654 seconds
+[2025-12-18 17:03:04,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499490 seconds
+[2025-12-18 17:03:04,814][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392718 seconds
+[2025-12-18 17:03:05,622][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979419 seconds
+[2025-12-18 17:03:05,728][openai._base_client][INFO] - Retrying request to /chat/completions in 0.914083 seconds
+[2025-12-18 17:03:10,212][openai._base_client][INFO] - Retrying request to /chat/completions in 0.992291 seconds
+[2025-12-18 17:03:10,316][openai._base_client][INFO] - Retrying request to /chat/completions in 0.778111 seconds
+[2025-12-18 17:03:26,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:03:31,072][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:03:33,978][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:03:37,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:03:37,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:03:40,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:03:43,779][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:03:48,789][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493400 seconds
+[2025-12-18 17:03:54,291][openai._base_client][INFO] - Retrying request to /chat/completions in 0.817178 seconds
+[2025-12-18 17:04:33,762][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:34,009][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496277 seconds
+[2025-12-18 17:04:37,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:38,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:38,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:38,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:39,009][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467064 seconds
+[2025-12-18 17:04:39,010][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436579 seconds
+[2025-12-18 17:04:39,011][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375909 seconds
+[2025-12-18 17:04:39,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401537 seconds
+[2025-12-18 17:04:39,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430583 seconds
+[2025-12-18 17:04:39,508][openai._base_client][INFO] - Retrying request to /chat/completions in 0.977641 seconds
+[2025-12-18 17:04:42,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:42,325][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:43,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:44,390][openai._base_client][INFO] - Retrying request to /chat/completions in 0.987516 seconds
+[2025-12-18 17:04:44,418][openai._base_client][INFO] - Retrying request to /chat/completions in 0.809916 seconds
+[2025-12-18 17:04:44,447][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781148 seconds
+[2025-12-18 17:04:44,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842126 seconds
+[2025-12-18 17:04:44,479][openai._base_client][INFO] - Retrying request to /chat/completions in 0.900376 seconds
+[2025-12-18 17:04:50,495][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436223 seconds
+[2025-12-18 17:04:54,433][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:04:55,243][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436003 seconds
+[2025-12-18 17:04:55,300][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380341 seconds
+[2025-12-18 17:04:55,384][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447155 seconds
+[2025-12-18 17:04:55,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.994873 seconds
+[2025-12-18 17:04:59,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:00,683][openai._base_client][INFO] - Retrying request to /chat/completions in 0.966163 seconds
+[2025-12-18 17:05:00,684][openai._base_client][INFO] - Retrying request to /chat/completions in 0.927851 seconds
+[2025-12-18 17:05:00,835][openai._base_client][INFO] - Retrying request to /chat/completions in 0.789612 seconds
+[2025-12-18 17:05:03,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:06,940][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392698 seconds
+[2025-12-18 17:05:08,054][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:09,602][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:11,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430225 seconds
+[2025-12-18 17:05:11,630][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477825 seconds
+[2025-12-18 17:05:11,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:12,336][openai._base_client][INFO] - Retrying request to /chat/completions in 0.770170 seconds
+[2025-12-18 17:05:14,592][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:14,929][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:17,113][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969658 seconds
+[2025-12-18 17:05:19,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:25,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:26,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:28,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:33,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:33,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:34,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:36,650][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:40,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:42,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:42,848][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:51,732][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:05:55,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:06,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:08,223][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:09,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:10,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:11,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:13,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:15,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:17,529][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:17,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:18,467][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:19,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:20,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:20,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:22,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:23,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:24,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:30,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:32,202][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:32,943][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:36,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:40,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:44,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:44,394][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:45,342][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:50,981][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:53,111][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:53,404][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:55,159][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:56,648][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:06:56,816][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:01,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:04,386][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:05,258][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:05,920][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:10,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:10,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:12,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:12,465][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:16,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:18,089][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:20,713][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:20,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:22,912][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:25,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:26,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:27,206][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:28,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:28,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:31,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:32,658][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:33,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:34,657][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:37,166][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:37,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:38,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:46,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:46,856][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:48,925][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:49,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:49,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:50,191][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:50,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:52,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:53,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:53,472][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:56,313][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:59,742][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:07:59,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:05,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:05,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:10,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:14,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:17,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:19,425][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:28,059][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:31,609][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:31,903][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:39,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:52,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:54,430][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:08:58,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:04,241][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449952 seconds
+[2025-12-18 17:09:04,452][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441041 seconds
+[2025-12-18 17:09:04,772][openai._base_client][INFO] - Retrying request to /chat/completions in 0.925756 seconds
+[2025-12-18 17:09:09,267][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485726 seconds
+[2025-12-18 17:09:09,268][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441569 seconds
+[2025-12-18 17:09:09,269][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449127 seconds
+[2025-12-18 17:09:09,270][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382284 seconds
+[2025-12-18 17:09:09,734][openai._base_client][INFO] - Retrying request to /chat/completions in 0.973687 seconds
+[2025-12-18 17:09:09,790][openai._base_client][INFO] - Retrying request to /chat/completions in 0.976577 seconds
+[2025-12-18 17:09:09,798][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812769 seconds
+[2025-12-18 17:09:09,836][openai._base_client][INFO] - Retrying request to /chat/completions in 0.909421 seconds
+[2025-12-18 17:09:09,895][openai._base_client][INFO] - Retrying request to /chat/completions in 0.874692 seconds
+[2025-12-18 17:09:10,867][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435402 seconds
+[2025-12-18 17:09:10,914][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478909 seconds
+[2025-12-18 17:09:11,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:15,700][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457506 seconds
+[2025-12-18 17:09:15,704][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399803 seconds
+[2025-12-18 17:09:15,853][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381875 seconds
+[2025-12-18 17:09:16,187][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860031 seconds
+[2025-12-18 17:09:16,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.981128 seconds
+[2025-12-18 17:09:16,308][openai._base_client][INFO] - Retrying request to /chat/completions in 0.775310 seconds
+[2025-12-18 17:09:16,316][openai._base_client][INFO] - Retrying request to /chat/completions in 0.938197 seconds
+[2025-12-18 17:09:16,396][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815749 seconds
+[2025-12-18 17:09:17,209][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393758 seconds
+[2025-12-18 17:09:17,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:20,777][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475396 seconds
+[2025-12-18 17:09:21,335][openai._base_client][INFO] - Retrying request to /chat/completions in 0.846024 seconds
+[2025-12-18 17:09:22,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:28,251][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:28,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:30,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:34,294][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:09:34,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:00,047][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:04,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:07,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:08,294][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:16,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:21,823][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473265 seconds
+[2025-12-18 17:10:21,927][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405958 seconds
+[2025-12-18 17:10:21,929][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384325 seconds
+[2025-12-18 17:10:21,929][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474958 seconds
+[2025-12-18 17:10:21,986][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391343 seconds
+[2025-12-18 17:10:22,002][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431755 seconds
+[2025-12-18 17:10:22,020][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443186 seconds
+[2025-12-18 17:10:22,378][openai._base_client][INFO] - Retrying request to /chat/completions in 0.811226 seconds
+[2025-12-18 17:10:22,412][openai._base_client][INFO] - Retrying request to /chat/completions in 0.809984 seconds
+[2025-12-18 17:10:22,460][openai._base_client][INFO] - Retrying request to /chat/completions in 0.885928 seconds
+[2025-12-18 17:10:22,484][openai._base_client][INFO] - Retrying request to /chat/completions in 0.929720 seconds
+[2025-12-18 17:10:22,541][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807601 seconds
+[2025-12-18 17:10:26,847][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430986 seconds
+[2025-12-18 17:10:27,626][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:30,824][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:33,549][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:34,328][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:35,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:37,557][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:40,919][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:45,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:47,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:51,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:10:54,906][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:11:17,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:11:20,593][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:11:25,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376766 seconds
+[2025-12-18 17:11:33,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:12:34,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:12:55,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:01,130][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445019 seconds
+[2025-12-18 17:13:29,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:29,567][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485441 seconds
+[2025-12-18 17:13:34,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415334 seconds
+[2025-12-18 17:13:34,574][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481541 seconds
+[2025-12-18 17:13:34,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395813 seconds
+[2025-12-18 17:13:35,057][openai._base_client][INFO] - Retrying request to /chat/completions in 0.750048 seconds
+[2025-12-18 17:13:35,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:37,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 503 Service Unavailable"
+[2025-12-18 17:13:37,375][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433531 seconds
+[2025-12-18 17:13:38,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:40,061][openai._base_client][INFO] - Retrying request to /chat/completions in 0.861315 seconds
+[2025-12-18 17:13:42,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:43,833][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:48,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:51,141][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:53,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:56,783][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:57,630][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:13:59,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:01,258][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:01,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:01,317][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:04,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:07,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:08,012][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:08,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:08,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:09,232][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:10,489][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:10,699][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:13,128][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:13,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:13,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:16,060][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:16,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:20,914][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:22,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:22,992][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:24,779][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:25,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:26,713][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:27,916][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:28,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:32,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:35,047][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:37,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:41,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:41,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:43,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:44,641][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:47,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:47,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:49,922][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:50,287][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:14:58,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:02,215][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:07,873][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:07,884][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:09,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:09,186][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:09,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:09,664][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:16,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:17,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:23,008][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:23,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:24,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:28,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:29,989][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:30,840][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:37,856][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:40,342][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:40,386][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:40,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:41,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:47,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:50,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:50,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:51,186][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:51,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:52,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:53,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:53,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:15:57,063][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:02,637][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:05,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:06,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:11,999][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:12,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:12,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:14,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:14,918][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:15,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:15,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:16,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:16,641][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:18,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:20,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:25,794][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:25,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:26,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:31,433][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:31,758][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:35,383][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:38,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:38,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:39,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:39,461][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:47,106][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:51,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:52,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:16:53,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:00,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:00,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:01,095][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:11,265][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:11,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:14,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:16,284][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:26,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:26,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:28,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:41,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:17:49,887][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:18:00,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:18:06,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:18:11,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 17:18:16,185][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462800 seconds
diff --git a/outputs/2025-12-18/17-18-36/.hydra/config.yaml b/outputs/2025-12-18/17-18-36/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a678c0dbcc52c29fbbb53b54a08bb279d975b0da
--- /dev/null
+++ b/outputs/2025-12-18/17-18-36/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/17-18-36/.hydra/hydra.yaml b/outputs/2025-12-18/17-18-36/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0dd28992b341a706fa3e4af13bd97c94525330ea
--- /dev/null
+++ b/outputs/2025-12-18/17-18-36/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/17-18-36
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/17-18-36/.hydra/overrides.yaml b/outputs/2025-12-18/17-18-36/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/17-18-36/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/17-18-36/eval_api.log b/outputs/2025-12-18/17-18-36/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..aec54705bfa83e4df47bb7455a84ebe9e1b8a06f
--- /dev/null
+++ b/outputs/2025-12-18/17-18-36/eval_api.log
@@ -0,0 +1,10 @@
+[2025-12-18 17:18:51,972][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f7abe697a30>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 44e4983c-15bc-447e-bdd1-73d2bf20144c)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 17:18:51,974][huggingface_hub.utils._http][WARNING] - Retrying in 1s [Retry 1/5].
+[2025-12-18 17:19:03,053][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f7abe6978e0>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 2bbcb865-113d-43af-affb-081d50ed6f03)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 17:19:03,054][huggingface_hub.utils._http][WARNING] - Retrying in 2s [Retry 2/5].
+[2025-12-18 17:19:15,070][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f7abe63dc30>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 1663ac25-2dfe-45bd-ba77-3db1c0be2a3c)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 17:19:15,071][huggingface_hub.utils._http][WARNING] - Retrying in 4s [Retry 3/5].
+[2025-12-18 17:19:29,156][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f7abe63da80>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 36a1ee5c-e01c-4751-a0aa-5443706e8072)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 17:19:29,157][huggingface_hub.utils._http][WARNING] - Retrying in 8s [Retry 4/5].
+[2025-12-18 17:19:47,242][huggingface_hub.utils._http][WARNING] - '(MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f7abe63e0e0>: Failed to establish a new connection: [Errno 101] Network is unreachable'))"), '(Request ID: 5a8eacf8-1b3a-4c5d-94af-3529e1dd6a8a)')' thrown while requesting HEAD https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/resolve/main/tokenizer_config.json
+[2025-12-18 17:19:47,243][huggingface_hub.utils._http][WARNING] - Retrying in 8s [Retry 5/5].
diff --git a/outputs/2025-12-18/17-20-05/.hydra/config.yaml b/outputs/2025-12-18/17-20-05/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7fdd5491946697bd70ba23106a366ca0281516b9
--- /dev/null
+++ b/outputs/2025-12-18/17-20-05/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/17-20-05/.hydra/hydra.yaml b/outputs/2025-12-18/17-20-05/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a8d987b56be2952dd05a70d3d824614e35a6f5ba
--- /dev/null
+++ b/outputs/2025-12-18/17-20-05/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/17-20-05
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/17-20-05/.hydra/overrides.yaml b/outputs/2025-12-18/17-20-05/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/17-20-05/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/18-22-00/.hydra/config.yaml b/outputs/2025-12-18/18-22-00/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48226dec8d605f285757e1345372a8e5657058b7
--- /dev/null
+++ b/outputs/2025-12-18/18-22-00/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: deepseek-v3
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/18-22-00/.hydra/hydra.yaml b/outputs/2025-12-18/18-22-00/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b099e408da23602d94cd39e6d86391f6394071a3
--- /dev/null
+++ b/outputs/2025-12-18/18-22-00/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/18-22-00
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/18-22-00/.hydra/overrides.yaml b/outputs/2025-12-18/18-22-00/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/18-22-00/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/18-22-00/eval_api.log b/outputs/2025-12-18/18-22-00/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/18-22-29/.hydra/config.yaml b/outputs/2025-12-18/18-22-29/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..713bb3d1a7f277115fd7c67cc9b91499ebe07948
--- /dev/null
+++ b/outputs/2025-12-18/18-22-29/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 128
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 128
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: ark-deepseek-v3-250324
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-18/18-22-29/.hydra/hydra.yaml b/outputs/2025-12-18/18-22-29/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..978da5a0bb36230522260570e08b4c142d4158c6
--- /dev/null
+++ b/outputs/2025-12-18/18-22-29/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/18-22-29
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/18-22-29/.hydra/overrides.yaml b/outputs/2025-12-18/18-22-29/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-18/18-22-29/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-18/18-22-29/eval_api.log b/outputs/2025-12-18/18-22-29/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..4edc92cd66dcaeb137949023b39914af3330c00e
--- /dev/null
+++ b/outputs/2025-12-18/18-22-29/eval_api.log
@@ -0,0 +1,753 @@
+[2025-12-18 18:22:35,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:35,900][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:36,340][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481818 seconds
+[2025-12-18 18:22:36,341][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478382 seconds
+[2025-12-18 18:22:36,341][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424171 seconds
+[2025-12-18 18:22:36,342][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401772 seconds
+[2025-12-18 18:22:36,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:37,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:37,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:38,313][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:38,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:38,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:38,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:38,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:39,161][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:39,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:40,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:41,747][openai._base_client][INFO] - Retrying request to /chat/completions in 0.942169 seconds
+[2025-12-18 18:22:41,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:41,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:41,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:42,091][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:42,523][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:42,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:42,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:43,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:43,630][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:43,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:44,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:44,479][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:45,688][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:45,809][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:45,940][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:45,993][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:47,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:47,109][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:47,136][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:47,222][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:48,267][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:48,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:49,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:49,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:50,509][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:51,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:51,134][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:51,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:51,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:51,428][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:52,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:52,931][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:53,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:53,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:53,508][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:53,734][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:53,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:54,573][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:54,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:55,928][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:56,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:56,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:56,944][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:57,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:57,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:57,323][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:57,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:57,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:58,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:58,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:58,573][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:58,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:22:58,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:00,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:00,860][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:00,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:01,268][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:01,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:01,828][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:01,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:02,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:02,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:02,982][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:03,151][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:03,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:03,816][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:04,128][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:04,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:04,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:04,900][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:05,091][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:05,151][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:05,700][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:05,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:06,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:06,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:06,325][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:06,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:07,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:08,813][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:09,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:09,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:10,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:10,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:10,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:10,693][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:10,716][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:11,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:11,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:12,012][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:12,363][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:12,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:12,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:13,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:13,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:14,056][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:14,109][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:15,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:16,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:16,749][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:16,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:16,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:16,992][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:17,134][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:18,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:18,149][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:18,250][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:18,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:18,693][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:19,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:19,248][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:19,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:19,912][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:22,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:27,091][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494757 seconds
+[2025-12-18 18:23:33,410][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:33,775][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458292 seconds
+[2025-12-18 18:23:37,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:38,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:38,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:38,776][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412427 seconds
+[2025-12-18 18:23:38,777][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397826 seconds
+[2025-12-18 18:23:38,777][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476722 seconds
+[2025-12-18 18:23:38,779][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480005 seconds
+[2025-12-18 18:23:39,236][openai._base_client][INFO] - Retrying request to /chat/completions in 0.836731 seconds
+[2025-12-18 18:23:39,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:40,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:40,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:40,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:40,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:41,109][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:41,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:42,668][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:42,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:43,290][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:43,303][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:43,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:44,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.896126 seconds
+[2025-12-18 18:23:44,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:44,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:44,760][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:45,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:45,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:46,424][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:46,429][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:46,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:46,843][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:47,329][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:47,620][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:47,940][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:48,189][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:48,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:48,378][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:48,573][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:48,916][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:49,737][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:50,271][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:50,325][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:50,519][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:51,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:51,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:52,251][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:52,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:52,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:52,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:52,968][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:53,949][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:54,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:54,752][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:54,756][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:55,163][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404748 seconds
+[2025-12-18 18:23:55,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:55,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:56,048][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:56,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:56,750][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:57,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:57,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:57,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:57,945][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:58,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:58,606][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:58,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:58,630][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:59,039][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:59,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:59,383][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:23:59,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:00,572][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816640 seconds
+[2025-12-18 18:24:00,623][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:01,294][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:01,671][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:02,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:02,470][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:02,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:02,825][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:02,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:03,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:03,148][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:03,253][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:04,419][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:04,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:04,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:04,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:05,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:05,522][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:05,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:05,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:06,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:06,620][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:07,046][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:07,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:07,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:07,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:08,287][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:08,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:09,364][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:10,020][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:10,148][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:10,267][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:10,750][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:11,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:11,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:11,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:11,749][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:12,092][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:12,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:12,659][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:12,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:13,271][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:13,405][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:13,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:13,848][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:14,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:15,090][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:15,437][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:15,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:15,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:16,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:16,702][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:16,978][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:17,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:17,329][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:17,405][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:18,013][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:18,383][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:18,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:19,351][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:19,517][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:19,654][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:21,768][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:26,778][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408189 seconds
+[2025-12-18 18:24:32,197][openai._base_client][INFO] - Retrying request to /chat/completions in 0.785745 seconds
+[2025-12-18 18:24:32,221][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:36,515][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:36,891][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424929 seconds
+[2025-12-18 18:24:40,724][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:41,577][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:41,714][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:41,895][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385410 seconds
+[2025-12-18 18:24:41,898][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497203 seconds
+[2025-12-18 18:24:41,898][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480661 seconds
+[2025-12-18 18:24:41,984][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:42,319][openai._base_client][INFO] - Retrying request to /chat/completions in 0.779380 seconds
+[2025-12-18 18:24:42,473][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:42,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:42,749][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:42,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:42,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:43,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:43,744][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:44,236][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:44,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:45,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:45,373][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:45,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:45,963][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:47,284][openai._base_client][INFO] - Retrying request to /chat/completions in 0.792881 seconds
+[2025-12-18 18:24:47,382][openai._base_client][INFO] - Retrying request to /chat/completions in 0.785477 seconds
+[2025-12-18 18:24:47,399][openai._base_client][INFO] - Retrying request to /chat/completions in 0.976042 seconds
+[2025-12-18 18:24:47,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:47,668][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:47,781][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:48,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:48,534][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:48,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:48,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:49,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:49,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:50,283][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:50,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:50,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:51,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:51,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:52,815][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:53,108][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401731 seconds
+[2025-12-18 18:24:53,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:53,803][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:54,127][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:54,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:54,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:55,408][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:55,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:56,383][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:56,529][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:57,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:57,187][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:58,086][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456077 seconds
+[2025-12-18 18:24:58,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:58,174][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466460 seconds
+[2025-12-18 18:24:58,378][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:58,383][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470648 seconds
+[2025-12-18 18:24:58,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:58,513][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781583 seconds
+[2025-12-18 18:24:58,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:59,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:24:59,623][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:00,007][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:00,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:00,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:01,141][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:01,292][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:01,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:02,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:02,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:02,925][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:03,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:03,546][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901673 seconds
+[2025-12-18 18:25:03,643][openai._base_client][INFO] - Retrying request to /chat/completions in 0.912307 seconds
+[2025-12-18 18:25:03,667][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:03,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:03,856][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815215 seconds
+[2025-12-18 18:25:04,410][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:04,568][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:05,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:05,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:06,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:06,349][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:06,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:07,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:07,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:07,280][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:07,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:08,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:08,798][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:09,301][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411307 seconds
+[2025-12-18 18:25:09,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:09,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:09,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:09,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:10,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:10,494][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:10,505][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:10,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:11,059][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:11,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:12,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:12,762][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:13,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:14,229][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:14,457][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483583 seconds
+[2025-12-18 18:25:14,716][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933038 seconds
+[2025-12-18 18:25:14,920][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:14,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:15,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:15,107][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:15,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:15,252][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:15,716][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:15,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:16,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:17,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:17,363][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:17,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:18,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:18,565][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:19,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:19,583][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:19,945][openai._base_client][INFO] - Retrying request to /chat/completions in 0.827115 seconds
+[2025-12-18 18:25:20,119][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:20,243][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:20,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:20,648][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:20,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:21,166][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:21,282][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:21,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:22,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:22,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:22,731][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:23,327][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:23,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:24,038][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:25,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:26,172][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:31,184][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429783 seconds
+[2025-12-18 18:25:35,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:35,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:36,200][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460702 seconds
+[2025-12-18 18:25:36,201][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441075 seconds
+[2025-12-18 18:25:36,202][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464995 seconds
+[2025-12-18 18:25:37,228][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:37,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:37,316][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:41,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:41,529][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:44,301][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:44,759][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479335 seconds
+[2025-12-18 18:25:44,760][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457442 seconds
+[2025-12-18 18:25:44,763][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497264 seconds
+[2025-12-18 18:25:49,765][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459946 seconds
+[2025-12-18 18:25:49,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427146 seconds
+[2025-12-18 18:25:49,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385703 seconds
+[2025-12-18 18:25:49,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468646 seconds
+[2025-12-18 18:25:49,768][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496872 seconds
+[2025-12-18 18:25:49,768][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493277 seconds
+[2025-12-18 18:25:49,768][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469997 seconds
+[2025-12-18 18:25:49,768][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388685 seconds
+[2025-12-18 18:25:49,769][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488737 seconds
+[2025-12-18 18:25:49,769][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422920 seconds
+[2025-12-18 18:25:49,769][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458008 seconds
+[2025-12-18 18:25:49,770][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482015 seconds
+[2025-12-18 18:25:49,770][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408911 seconds
+[2025-12-18 18:25:50,241][openai._base_client][INFO] - Retrying request to /chat/completions in 0.895637 seconds
+[2025-12-18 18:25:50,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:50,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:54,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:54,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:55,156][openai._base_client][INFO] - Retrying request to /chat/completions in 0.945902 seconds
+[2025-12-18 18:25:55,159][openai._base_client][INFO] - Retrying request to /chat/completions in 0.952721 seconds
+[2025-12-18 18:25:55,182][openai._base_client][INFO] - Retrying request to /chat/completions in 0.973721 seconds
+[2025-12-18 18:25:55,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969631 seconds
+[2025-12-18 18:25:55,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905406 seconds
+[2025-12-18 18:25:55,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.785999 seconds
+[2025-12-18 18:25:55,255][openai._base_client][INFO] - Retrying request to /chat/completions in 0.764069 seconds
+[2025-12-18 18:25:55,267][openai._base_client][INFO] - Retrying request to /chat/completions in 0.989344 seconds
+[2025-12-18 18:25:55,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:55,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:55,768][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:56,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:56,207][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:57,873][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:58,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:59,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:59,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:25:59,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:00,997][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:01,052][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:01,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:01,711][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:01,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:02,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:02,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:03,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:04,293][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:04,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:04,624][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:05,170][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:05,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:06,028][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472117 seconds
+[2025-12-18 18:26:06,030][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457788 seconds
+[2025-12-18 18:26:06,117][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376019 seconds
+[2025-12-18 18:26:06,143][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380330 seconds
+[2025-12-18 18:26:06,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:06,262][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378674 seconds
+[2025-12-18 18:26:06,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:06,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:07,063][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:08,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:08,161][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:08,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:09,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:09,842][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:10,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:11,470][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:11,491][openai._base_client][INFO] - Retrying request to /chat/completions in 0.822104 seconds
+[2025-12-18 18:26:11,496][openai._base_client][INFO] - Retrying request to /chat/completions in 0.875331 seconds
+[2025-12-18 18:26:11,503][openai._base_client][INFO] - Retrying request to /chat/completions in 0.958750 seconds
+[2025-12-18 18:26:11,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:11,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:12,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:12,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:12,840][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:12,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:13,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:13,945][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:14,108][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:14,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:14,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:14,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:15,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:16,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:16,511][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:16,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:17,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:17,202][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:17,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:17,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:18,059][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:18,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:18,490][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:18,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:18,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:19,063][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:19,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:20,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:20,323][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:20,349][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:20,493][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:20,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:21,195][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:21,202][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:21,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:21,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:22,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:22,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:23,035][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:23,076][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:23,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:23,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:23,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:23,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:24,988][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:25,113][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:25,148][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:25,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:26,136][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:26,283][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:26,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:27,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:27,640][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:27,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:28,005][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:28,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:28,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:28,917][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:28,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:29,272][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:29,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:29,910][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:30,430][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:30,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:30,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:31,144][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:31,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:32,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:32,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:32,857][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:33,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:33,518][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:33,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:33,912][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:34,038][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:34,124][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:34,602][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:35,039][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:35,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:35,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:36,848][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:36,854][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:37,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:42,721][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446733 seconds
+[2025-12-18 18:26:47,731][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462074 seconds
+[2025-12-18 18:26:47,732][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443588 seconds
+[2025-12-18 18:26:47,732][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492333 seconds
+[2025-12-18 18:26:48,173][openai._base_client][INFO] - Retrying request to /chat/completions in 0.996744 seconds
+[2025-12-18 18:26:49,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:53,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:53,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:54,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:55,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:26:56,172][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456002 seconds
+[2025-12-18 18:26:56,174][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461670 seconds
+[2025-12-18 18:26:56,176][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400318 seconds
+[2025-12-18 18:26:56,178][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447467 seconds
+[2025-12-18 18:27:00,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:01,123][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:01,177][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435281 seconds
+[2025-12-18 18:27:01,179][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383870 seconds
+[2025-12-18 18:27:01,179][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486374 seconds
+[2025-12-18 18:27:01,181][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447627 seconds
+[2025-12-18 18:27:01,181][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424974 seconds
+[2025-12-18 18:27:01,399][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:01,439][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:01,579][openai._base_client][INFO] - Retrying request to /chat/completions in 0.965930 seconds
+[2025-12-18 18:27:01,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.784952 seconds
+[2025-12-18 18:27:01,632][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803357 seconds
+[2025-12-18 18:27:01,638][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860606 seconds
+[2025-12-18 18:27:02,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:02,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:02,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:04,477][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:05,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:05,814][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:06,567][openai._base_client][INFO] - Retrying request to /chat/completions in 0.825216 seconds
+[2025-12-18 18:27:06,608][openai._base_client][INFO] - Retrying request to /chat/completions in 0.790139 seconds
+[2025-12-18 18:27:06,616][openai._base_client][INFO] - Retrying request to /chat/completions in 0.852893 seconds
+[2025-12-18 18:27:06,632][openai._base_client][INFO] - Retrying request to /chat/completions in 0.984308 seconds
+[2025-12-18 18:27:06,670][openai._base_client][INFO] - Retrying request to /chat/completions in 0.915046 seconds
+[2025-12-18 18:27:06,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:07,788][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:08,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:08,291][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:09,217][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:09,241][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:09,704][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:09,982][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:11,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:11,768][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:12,421][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382972 seconds
+[2025-12-18 18:27:12,442][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382719 seconds
+[2025-12-18 18:27:12,465][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:12,506][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457216 seconds
+[2025-12-18 18:27:12,553][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496436 seconds
+[2025-12-18 18:27:13,375][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:13,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:14,368][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:15,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:16,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:16,992][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:17,327][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:17,399][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385509 seconds
+[2025-12-18 18:27:17,405][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420035 seconds
+[2025-12-18 18:27:17,475][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432317 seconds
+[2025-12-18 18:27:17,592][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375139 seconds
+[2025-12-18 18:27:17,622][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387863 seconds
+[2025-12-18 18:27:17,807][openai._base_client][INFO] - Retrying request to /chat/completions in 0.794511 seconds
+[2025-12-18 18:27:17,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.841897 seconds
+[2025-12-18 18:27:17,967][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781064 seconds
+[2025-12-18 18:27:18,052][openai._base_client][INFO] - Retrying request to /chat/completions in 0.823051 seconds
+[2025-12-18 18:27:19,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:19,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:20,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:20,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:21,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:22,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:22,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.847480 seconds
+[2025-12-18 18:27:22,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.959366 seconds
+[2025-12-18 18:27:22,855][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:22,910][openai._base_client][INFO] - Retrying request to /chat/completions in 0.825026 seconds
+[2025-12-18 18:27:22,971][openai._base_client][INFO] - Retrying request to /chat/completions in 0.769508 seconds
+[2025-12-18 18:27:23,014][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967806 seconds
+[2025-12-18 18:27:23,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:24,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:25,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:25,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:25,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:26,180][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:27,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:27,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:27,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:28,612][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417495 seconds
+[2025-12-18 18:27:28,883][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:28,988][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:29,217][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:29,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:29,403][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:29,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:29,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:30,372][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:30,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:31,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:31,982][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:32,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:32,192][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:32,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:32,676][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:32,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:33,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:33,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:33,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:34,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:34,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:34,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:34,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:34,639][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:35,623][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:35,929][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:35,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:36,133][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:36,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:36,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:36,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:36,471][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:36,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:37,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:37,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:37,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:38,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:39,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:39,699][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:39,744][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:39,906][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:40,299][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:40,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:40,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:40,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:41,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:41,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:41,217][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:41,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:41,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:42,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:42,943][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:44,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:44,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:44,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:45,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:45,103][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:45,186][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:45,313][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:45,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:45,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:45,922][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:46,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:46,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:47,076][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:47,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:47,815][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:52,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462191 seconds
+[2025-12-18 18:27:56,803][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:57,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:57,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:57,595][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:58,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:58,161][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:58,242][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:58,244][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:58,389][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:58,943][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:59,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:59,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:59,668][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:27:59,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-18 18:28:00,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-18/19-01-24/.hydra/config.yaml b/outputs/2025-12-18/19-01-24/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0000d28b8ee6f7ccc92dbf37f6a2df0588c0e482
--- /dev/null
+++ b/outputs/2025-12-18/19-01-24/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1-starpos-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-18/19-01-24/.hydra/hydra.yaml b/outputs/2025-12-18/19-01-24/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..46c70578071826069dc849e9722e63e702086c0f
--- /dev/null
+++ b/outputs/2025-12-18/19-01-24/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-1-starpos-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/19-01-24
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/19-01-24/.hydra/overrides.yaml b/outputs/2025-12-18/19-01-24/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c84231cbb0be8ce6115bfd65531aedd2c0413fea
--- /dev/null
+++ b/outputs/2025-12-18/19-01-24/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- trainer.experiment_name=rubikscube-1-starpos-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-18/19-01-24/train.log b/outputs/2025-12-18/19-01-24/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/20-09-55/.hydra/hydra.yaml b/outputs/2025-12-18/20-09-55/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d8a041c0ac0a5fb9b49260cd2c4af0e547f29531
--- /dev/null
+++ b/outputs/2025-12-18/20-09-55/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube1_1218
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos-frommlp-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube1_1218,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-1-starpos-frommlp-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/20-09-55
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/20-09-55/.hydra/overrides.yaml b/outputs/2025-12-18/20-09-55/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..23de3b3f74c8328039fe42978c9abf39a8b3fb80
--- /dev/null
+++ b/outputs/2025-12-18/20-09-55/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube1_1218
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- trainer.experiment_name=rubikscube-1-starpos-frommlp-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-18/20-09-55/train.log b/outputs/2025-12-18/20-09-55/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/20-30-46/.hydra/config.yaml b/outputs/2025-12-18/20-30-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..51fe22bdf90cce1bf9327e2b3e71dbff7a2bc85c
--- /dev/null
+++ b/outputs/2025-12-18/20-30-46/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-frommlp-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1218
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-18/20-30-46/.hydra/hydra.yaml b/outputs/2025-12-18/20-30-46/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..234abab3a541f850798ccb0be3a59011392b1149
--- /dev/null
+++ b/outputs/2025-12-18/20-30-46/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1218
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1218,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/20-30-46
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/20-30-46/.hydra/overrides.yaml b/outputs/2025-12-18/20-30-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0305adac9bc4356cc1e54ffa2bd1dfb163ea63fd
--- /dev/null
+++ b/outputs/2025-12-18/20-30-46/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1218
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-18/20-30-46/train.log b/outputs/2025-12-18/20-30-46/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/20-58-58/.hydra/config.yaml b/outputs/2025-12-18/20-58-58/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7118afea34d09153ea708460685b9e6e63cea68a
--- /dev/null
+++ b/outputs/2025-12-18/20-58-58/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-1-starpos-frommlp-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube1_1218
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-18/20-58-58/train.log b/outputs/2025-12-18/20-58-58/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/21-11-13/.hydra/overrides.yaml b/outputs/2025-12-18/21-11-13/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5d5c4f435d4f4e53faf9458529643644e5433a9c
--- /dev/null
+++ b/outputs/2025-12-18/21-11-13/.hydra/overrides.yaml
@@ -0,0 +1,20 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_SPA_frozenlake
+- custom_envs.CoordFrozenLake.env_instruction='You are solving the FrozenLake puzzle.\nForbid
+  the whole and go to the target.\nYou may move to the unintended direction due to
+  the slippery ice.\nExample answer format:\n<think>To forbid the hole and go to the
+  target, I should go left then go up.</think>\n<answer>Left || Up</answer>\nA sample
+  full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+  at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+  2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now on
+  G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+  || Left || Left</answer>'
+- trainer.experiment_name=frozenlake-static-starpos-SPA-7200
+- trainer.total_training_steps=1000
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-18/21-30-49/.hydra/config.yaml b/outputs/2025-12-18/21-30-49/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bf4a18c0538ea9582216e1a9717c22dc63bf308b
--- /dev/null
+++ b/outputs/2025-12-18/21-30-49/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1218
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-18/21-30-49/.hydra/overrides.yaml b/outputs/2025-12-18/21-30-49/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dee523f8162027ff87f8eb24d2e507f3063af1f8
--- /dev/null
+++ b/outputs/2025-12-18/21-30-49/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1218
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-18/21-30-49/train.log b/outputs/2025-12-18/21-30-49/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/21-46-47/.hydra/config.yaml b/outputs/2025-12-18/21-46-47/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ca0ed88858c03c2c0059b580be3e74df15c5bc4b
--- /dev/null
+++ b/outputs/2025-12-18/21-46-47/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-18/21-46-47/.hydra/hydra.yaml b/outputs/2025-12-18/21-46-47/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d041c961481a769cfae122a598c5385cb5d9a213
--- /dev/null
+++ b/outputs/2025-12-18/21-46-47/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/21-46-47
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/22-19-43/eval_api.log b/outputs/2025-12-18/22-19-43/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-18/22-23-34/.hydra/hydra.yaml b/outputs/2025-12-18/22-23-34/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..929bd8c4a3830f6dc6eca712d39a094117b0306f
--- /dev/null
+++ b/outputs/2025-12-18/22-23-34/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-18/22-23-34
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-18/22-34-55/train.log b/outputs/2025-12-18/22-34-55/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/01-11-07/.hydra/config.yaml b/outputs/2025-12-19/01-11-07/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..73037f0b6a989675a6b7f98f11a4c57f75c17950
--- /dev/null
+++ b/outputs/2025-12-19/01-11-07/.hydra/config.yaml
@@ -0,0 +1,358 @@
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+enable_response_mask: true
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_2
+lora:
+  rank: 0
+  alpha: 64
+  target_modules: all-linear
+actor_rollout_ref:
+  model:
+    path: ${model_path}
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  rollout:
+    name: vllm
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    tensor_model_parallel_size: 8
+    dtype: bfloat16
+    max_model_len: 7200
+    prompt_length: 1
+    response_length: 400
+    gpu_memory_utilization: 0.9
+    max_num_batched_tokens: 8192
+    enforce_eager: true
+    free_cache_engine: true
+    enable_chunked_prefill: false
+    disable_log_stats: false
+    do_sample: true
+    val_kwargs:
+      do_sample: true
+      temperature: 0.5
+      top_p: 1.0
+      top_k: -1
+      logprobs: 20
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 128
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+output:
+  dir: results/eval
+  filename: val_rollouts.pkl
+  append_timestamp: true
+  keep_batch_keys: null
+  keep_non_tensor_keys: null
+  keep_meta_info: true
diff --git a/outputs/2025-12-19/01-14-37/.hydra/config.yaml b/outputs/2025-12-19/01-14-37/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ece3e855a35400e3ed7b726cf8bd10deeb857d44
--- /dev/null
+++ b/outputs/2025-12-19/01-14-37/.hydra/config.yaml
@@ -0,0 +1,358 @@
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+enable_response_mask: true
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_2
+lora:
+  rank: 0
+  alpha: 64
+  target_modules: all-linear
+actor_rollout_ref:
+  model:
+    path: ${model_path}
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  rollout:
+    name: vllm
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    tensor_model_parallel_size: 8
+    dtype: bfloat16
+    max_model_len: 7200
+    prompt_length: 1
+    response_length: 400
+    gpu_memory_utilization: 0.9
+    max_num_batched_tokens: 8192
+    enforce_eager: true
+    free_cache_engine: true
+    enable_chunked_prefill: false
+    disable_log_stats: false
+    do_sample: true
+    val_kwargs:
+      do_sample: true
+      temperature: 0.5
+      top_p: 1.0
+      top_k: -1
+      logprobs: 20
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 128
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+output:
+  dir: results/eval
+  filename: val_rollouts.pkl
+  append_timestamp: true
+  keep_batch_keys: null
+  keep_non_tensor_keys: null
+  keep_meta_info: true
diff --git a/outputs/2025-12-19/01-15-10/.hydra/config.yaml b/outputs/2025-12-19/01-15-10/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0cbbd53d4d5203bb870074b754b9159efd0ef9dd
--- /dev/null
+++ b/outputs/2025-12-19/01-15-10/.hydra/config.yaml
@@ -0,0 +1,358 @@
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+enable_response_mask: true
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube_2
+lora:
+  rank: 0
+  alpha: 64
+  target_modules: all-linear
+actor_rollout_ref:
+  model:
+    path: ${model_path}
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  rollout:
+    name: vllm
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    tensor_model_parallel_size: 1
+    dtype: bfloat16
+    max_model_len: 7200
+    prompt_length: 1
+    response_length: 400
+    gpu_memory_utilization: 0.9
+    max_num_batched_tokens: 8192
+    enforce_eager: true
+    free_cache_engine: true
+    enable_chunked_prefill: false
+    disable_log_stats: false
+    do_sample: true
+    val_kwargs:
+      do_sample: true
+      temperature: 0.5
+      top_p: 1.0
+      top_k: -1
+      logprobs: 20
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 128
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+output:
+  dir: results/eval
+  filename: val_rollouts.pkl
+  append_timestamp: true
+  keep_batch_keys: null
+  keep_non_tensor_keys: null
+  keep_meta_info: true
diff --git a/outputs/2025-12-19/01-15-10/.hydra/overrides.yaml b/outputs/2025-12-19/01-15-10/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-19/01-15-10/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-19/01-20-00/.hydra/hydra.yaml b/outputs/2025-12-19/01-20-00/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f4a719b962bdbe05171a68dc830c837b5bdee6d
--- /dev/null
+++ b/outputs/2025-12-19/01-20-00/.hydra/hydra.yaml
@@ -0,0 +1,157 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: agent_proxy
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: eval.yaml
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/01-20-00
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/01-20-00/.hydra/overrides.yaml b/outputs/2025-12-19/01-20-00/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-19/01-20-00/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-19/01-59-08/.hydra/config.yaml b/outputs/2025-12-19/01-59-08/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4045c06e0c3e51bbfb10805e82e56722ed1a46ef
--- /dev/null
+++ b/outputs/2025-12-19/01-59-08/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-1-starpos-frommlp-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube1_1218
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/01-59-08/.hydra/hydra.yaml b/outputs/2025-12-19/01-59-08/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d96c538cb64bd732c3229fd8309e1bf1d5f6da56
--- /dev/null
+++ b/outputs/2025-12-19/01-59-08/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube1_1218
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - trainer.experiment_name=rubikscube-1-starpos-frommlp-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube1_1218,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-1-starpos-frommlp-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/01-59-08
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/02-26-30/.hydra/overrides.yaml b/outputs/2025-12-19/02-26-30/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a17a00e62e8c2ad0fdfa835ef662db2afc3eaf61
--- /dev/null
+++ b/outputs/2025-12-19/02-26-30/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/03-47-21/.hydra/config.yaml b/outputs/2025-12-19/03-47-21/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..28709107428813d4dba91d0a94cac280c6dce31e
--- /dev/null
+++ b/outputs/2025-12-19/03-47-21/.hydra/config.yaml
@@ -0,0 +1,853 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 1000
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: frozenlake-static-starpos-SPA-7200
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: You are solving the FrozenLake puzzle.\nForbid the whole and
+      go to the target.\nYou may move to the unintended direction due to the slippery
+      ice.\nExample answer format:\n<think>To forbid the hole and go to the target,
+      I should go left then go up.</think>\n<answer>Left || Up</answer>\nA sample
+      full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+      at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+      2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now
+      on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+      || Left || Left</answer>
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_SPA_frozenlake
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/03-47-21/.hydra/hydra.yaml b/outputs/2025-12-19/03-47-21/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..af4c22a36185bd1cd8bd1f391aba3f5519bcf73e
--- /dev/null
+++ b/outputs/2025-12-19/03-47-21/.hydra/hydra.yaml
@@ -0,0 +1,197 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_SPA_frozenlake
+    - custom_envs.CoordFrozenLake.env_instruction='You are solving the FrozenLake
+      puzzle.\nForbid the whole and go to the target.\nYou may move to the unintended
+      direction due to the slippery ice.\nExample answer format:\n<think>To forbid
+      the hole and go to the target, I should go left then go up.</think>\n<answer>Left
+      || Up</answer>\nA sample full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+      at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+      2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now
+      on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+      || Left || Left</answer>'
+    - trainer.experiment_name=frozenlake-static-starpos-SPA-7200
+    - trainer.total_training_steps=1000
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.CoordFrozenLake.env_instruction='You
+      are solving the FrozenLake puzzle.\nForbid the whole and go to the target.\nYou
+      may move to the unintended direction due to the slippery ice.\nExample answer
+      format:\n<think>To forbid the hole and go to the target, I should go left then
+      go up.</think>\n<answer>Left || Up</answer>\nA sample full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+      at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+      2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now
+      on G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+      || Left || Left</answer>',model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_SPA_frozenlake,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=frozenlake-static-starpos-SPA-7200,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it,trainer.total_training_steps=1000
+    id: ???
+    num: ???
+    config_name: _3_frozen_lake
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/03-47-21
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/03-47-21/.hydra/overrides.yaml b/outputs/2025-12-19/03-47-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6cafd9e302f5c7c7dda18780a8ee436f8a4de392
--- /dev/null
+++ b/outputs/2025-12-19/03-47-21/.hydra/overrides.yaml
@@ -0,0 +1,20 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_SPA_frozenlake
+- custom_envs.CoordFrozenLake.env_instruction='You are solving the FrozenLake puzzle.\nForbid
+  the whole and go to the target.\nYou may move to the unintended direction due to
+  the slippery ice.\nExample answer format:\n<think>To forbid the hole and go to the
+  target, I should go left then go up.</think>\n<answer>Left || Up</answer>\nA sample
+  full output is as follows:\n<think>\n<observation>\n_O__\nO___\nG___\n__P_\n</observation>\nPlayer
+  at (3,2); holes at (0,1) and (1,0); goal at (2,0). 1 Up move to safe ice\n(2,2).
+  2 Left slide to (2,1), adjacent to goal. 3 Left reach goal (2,0);\nplayer now on
+  G.\n<prediction>\n_O__\nO___\n√\n___\n____\n</prediction>\n</think> \n<answer>Up
+  || Left || Left</answer>'
+- trainer.experiment_name=frozenlake-static-starpos-SPA-7200
+- trainer.total_training_steps=1000
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/04-28-39/.hydra/config.yaml b/outputs/2025-12-19/04-28-39/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8d85af19c6313d67ca5c3805363ac68f4b36851e
--- /dev/null
+++ b/outputs/2025-12-19/04-28-39/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1218
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/04-28-39/.hydra/hydra.yaml b/outputs/2025-12-19/04-28-39/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..87c54bb05196aed4eaf328f7ad3960bda09fc27b
--- /dev/null
+++ b/outputs/2025-12-19/04-28-39/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1218
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1218,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/04-28-39
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/04-28-39/.hydra/overrides.yaml b/outputs/2025-12-19/04-28-39/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a8dd21ed0b757945a31803fe274c8cbb10d83d6d
--- /dev/null
+++ b/outputs/2025-12-19/04-28-39/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1218
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/05-13-22/.hydra/config.yaml b/outputs/2025-12-19/05-13-22/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..257f414e76e15041c72c3d1d5d8aa302e2af28ae
--- /dev/null
+++ b/outputs/2025-12-19/05-13-22/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/05-13-22/.hydra/overrides.yaml b/outputs/2025-12-19/05-13-22/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7a1657a1d6cfef73d2ef6e10b9a7ebcfb62432c8
--- /dev/null
+++ b/outputs/2025-12-19/05-13-22/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/05-13-22/train.log b/outputs/2025-12-19/05-13-22/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/05-36-41/.hydra/config.yaml b/outputs/2025-12-19/05-36-41/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..868580613a1f1f2acf78ffb7aa5beeca1d8c4ecf
--- /dev/null
+++ b/outputs/2025-12-19/05-36-41/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/05-36-41/.hydra/hydra.yaml b/outputs/2025-12-19/05-36-41/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b367d52ec271b2f5fd9a7e0c27760295232ca41a
--- /dev/null
+++ b/outputs/2025-12-19/05-36-41/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/05-36-41
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/05-36-41/.hydra/overrides.yaml b/outputs/2025-12-19/05-36-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bd7895d1958413da51884205faee7a51fcab294e
--- /dev/null
+++ b/outputs/2025-12-19/05-36-41/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/05-36-41/train.log b/outputs/2025-12-19/05-36-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/06-35-08/.hydra/config.yaml b/outputs/2025-12-19/06-35-08/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d3e7dbb045a31473cfaa28f42f69d6ea9b4aa33a
--- /dev/null
+++ b/outputs/2025-12-19/06-35-08/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-frommlp-1218
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1218
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/06-35-08/.hydra/hydra.yaml b/outputs/2025-12-19/06-35-08/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..19500f5ac3e14c2143090c2b42f13104639d6c4c
--- /dev/null
+++ b/outputs/2025-12-19/06-35-08/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1218
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1218
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1218,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1218,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/06-35-08
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/06-35-08/.hydra/overrides.yaml b/outputs/2025-12-19/06-35-08/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..376e356b620787304ac86e0915dbbfbba47a50de
--- /dev/null
+++ b/outputs/2025-12-19/06-35-08/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1218
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1218
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/09-24-20/.hydra/config.yaml b/outputs/2025-12-19/09-24-20/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3d9d8024064b0c1111d06b72fdbfe1034e699149
--- /dev/null
+++ b/outputs/2025-12-19/09-24-20/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/09-24-20/.hydra/hydra.yaml b/outputs/2025-12-19/09-24-20/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3fe854ec774647d36f6e8c9cb68f6b0c7c15fdee
--- /dev/null
+++ b/outputs/2025-12-19/09-24-20/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/09-24-20
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/09-24-20/.hydra/overrides.yaml b/outputs/2025-12-19/09-24-20/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..207405e27eca696ef1eb29501eeee9d723326bf1
--- /dev/null
+++ b/outputs/2025-12-19/09-24-20/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/09-24-20/train.log b/outputs/2025-12-19/09-24-20/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/10-46-22/.hydra/config.yaml b/outputs/2025-12-19/10-46-22/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fbfd2fc28c807d6d53f97e714bee1173f07cc4e6
--- /dev/null
+++ b/outputs/2025-12-19/10-46-22/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/10-46-22/.hydra/hydra.yaml b/outputs/2025-12-19/10-46-22/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..596c62f43bcb9acca75593d84d13cfdabfb357c0
--- /dev/null
+++ b/outputs/2025-12-19/10-46-22/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/10-46-22
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/10-46-22/.hydra/overrides.yaml b/outputs/2025-12-19/10-46-22/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c05022799344824b9ea185882e3be8dd53afc7ac
--- /dev/null
+++ b/outputs/2025-12-19/10-46-22/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/10-46-22/train.log b/outputs/2025-12-19/10-46-22/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/11-39-52/.hydra/config.yaml b/outputs/2025-12-19/11-39-52/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d22df20753bb32026270d35bba4a53081322ddaf
--- /dev/null
+++ b/outputs/2025-12-19/11-39-52/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/11-39-52/.hydra/hydra.yaml b/outputs/2025-12-19/11-39-52/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..42edcec91769e5b3131fd0defab8cea805c0c0f4
--- /dev/null
+++ b/outputs/2025-12-19/11-39-52/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/11-39-52
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/11-39-52/.hydra/overrides.yaml b/outputs/2025-12-19/11-39-52/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..973c57560a0c08b7718bb9f2d3fdec493a91fba1
--- /dev/null
+++ b/outputs/2025-12-19/11-39-52/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/11-39-52/train.log b/outputs/2025-12-19/11-39-52/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/12-15-22/.hydra/config.yaml b/outputs/2025-12-19/12-15-22/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48c8430509e5a41680193d5a12f3ef141786ebd1
--- /dev/null
+++ b/outputs/2025-12-19/12-15-22/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/12-15-22/.hydra/hydra.yaml b/outputs/2025-12-19/12-15-22/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bcc4f259c56188a1d4b981b90d7487f6b792a5cd
--- /dev/null
+++ b/outputs/2025-12-19/12-15-22/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/12-15-22
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/12-15-22/.hydra/overrides.yaml b/outputs/2025-12-19/12-15-22/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6c9212fc781154c859e8d4465297d55759321cb1
--- /dev/null
+++ b/outputs/2025-12-19/12-15-22/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/12-15-22/train.log b/outputs/2025-12-19/12-15-22/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/12-36-26/.hydra/config.yaml b/outputs/2025-12-19/12-36-26/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f91cadf732fb7b0e19e59891a5fee879b20ba22
--- /dev/null
+++ b/outputs/2025-12-19/12-36-26/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/12-36-26/.hydra/hydra.yaml b/outputs/2025-12-19/12-36-26/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f55a1dafa516ce6f713442d7419d5aa1eebaff55
--- /dev/null
+++ b/outputs/2025-12-19/12-36-26/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/12-36-26
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/12-36-26/.hydra/overrides.yaml b/outputs/2025-12-19/12-36-26/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..07ce9c8e8f92472cfdd0f977a36e421ebe1d48d1
--- /dev/null
+++ b/outputs/2025-12-19/12-36-26/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/12-36-26/train.log b/outputs/2025-12-19/12-36-26/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/12-52-21/.hydra/config.yaml b/outputs/2025-12-19/12-52-21/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f91cadf732fb7b0e19e59891a5fee879b20ba22
--- /dev/null
+++ b/outputs/2025-12-19/12-52-21/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/12-52-21/.hydra/hydra.yaml b/outputs/2025-12-19/12-52-21/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8230a14b9eaf549a0db60710e81fcd2bec652f2e
--- /dev/null
+++ b/outputs/2025-12-19/12-52-21/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/12-52-21
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/12-52-21/.hydra/overrides.yaml b/outputs/2025-12-19/12-52-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..07ce9c8e8f92472cfdd0f977a36e421ebe1d48d1
--- /dev/null
+++ b/outputs/2025-12-19/12-52-21/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/12-52-21/train.log b/outputs/2025-12-19/12-52-21/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/12-59-31/.hydra/config.yaml b/outputs/2025-12-19/12-59-31/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8f91cadf732fb7b0e19e59891a5fee879b20ba22
--- /dev/null
+++ b/outputs/2025-12-19/12-59-31/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/12-59-31/.hydra/hydra.yaml b/outputs/2025-12-19/12-59-31/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b12e49aff37af675db32f65b53922cd374e62dba
--- /dev/null
+++ b/outputs/2025-12-19/12-59-31/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=1
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=1,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/12-59-31
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/12-59-31/.hydra/overrides.yaml b/outputs/2025-12-19/12-59-31/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..07ce9c8e8f92472cfdd0f977a36e421ebe1d48d1
--- /dev/null
+++ b/outputs/2025-12-19/12-59-31/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=1
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/12-59-31/train.log b/outputs/2025-12-19/12-59-31/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/13-23-25/.hydra/config.yaml b/outputs/2025-12-19/13-23-25/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d22df20753bb32026270d35bba4a53081322ddaf
--- /dev/null
+++ b/outputs/2025-12-19/13-23-25/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/13-23-25/.hydra/hydra.yaml b/outputs/2025-12-19/13-23-25/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5eeda28065eb3b784df39c039c6f197e0d192861
--- /dev/null
+++ b/outputs/2025-12-19/13-23-25/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/13-23-25
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/13-23-25/.hydra/overrides.yaml b/outputs/2025-12-19/13-23-25/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..973c57560a0c08b7718bb9f2d3fdec493a91fba1
--- /dev/null
+++ b/outputs/2025-12-19/13-23-25/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/13-23-25/train.log b/outputs/2025-12-19/13-23-25/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/13-26-10/.hydra/config.yaml b/outputs/2025-12-19/13-26-10/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d22df20753bb32026270d35bba4a53081322ddaf
--- /dev/null
+++ b/outputs/2025-12-19/13-26-10/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/13-26-10/.hydra/hydra.yaml b/outputs/2025-12-19/13-26-10/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2044c511f3e3582390362d06a2fb76a812a41a02
--- /dev/null
+++ b/outputs/2025-12-19/13-26-10/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/13-26-10
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/13-26-10/.hydra/overrides.yaml b/outputs/2025-12-19/13-26-10/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..973c57560a0c08b7718bb9f2d3fdec493a91fba1
--- /dev/null
+++ b/outputs/2025-12-19/13-26-10/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/13-26-10/train.log b/outputs/2025-12-19/13-26-10/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/14-03-24/.hydra/config.yaml b/outputs/2025-12-19/14-03-24/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..48c8430509e5a41680193d5a12f3ef141786ebd1
--- /dev/null
+++ b/outputs/2025-12-19/14-03-24/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/14-03-24/.hydra/hydra.yaml b/outputs/2025-12-19/14-03-24/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f1411a64d6e2fcdac2da5a0c2c086cdc6800605d
--- /dev/null
+++ b/outputs/2025-12-19/14-03-24/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/14-03-24
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/14-03-24/.hydra/overrides.yaml b/outputs/2025-12-19/14-03-24/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6c9212fc781154c859e8d4465297d55759321cb1
--- /dev/null
+++ b/outputs/2025-12-19/14-03-24/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/14-03-24/train.log b/outputs/2025-12-19/14-03-24/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/15-22-31/.hydra/config.yaml b/outputs/2025-12-19/15-22-31/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..564d99e5967016291b94fa5fbbac71d9c2937381
--- /dev/null
+++ b/outputs/2025-12-19/15-22-31/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219_turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/15-22-31/.hydra/hydra.yaml b/outputs/2025-12-19/15-22-31/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..45c61f7ebecbe334e5c78fee942c547e3c27d69a
--- /dev/null
+++ b/outputs/2025-12-19/15-22-31/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219_turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219_turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/15-22-31
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/15-22-31/.hydra/overrides.yaml b/outputs/2025-12-19/15-22-31/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a2fc7305eb45b5e771043dd538343739e0c61eab
--- /dev/null
+++ b/outputs/2025-12-19/15-22-31/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219_turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/15-22-31/train.log b/outputs/2025-12-19/15-22-31/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/16-03-52/.hydra/config.yaml b/outputs/2025-12-19/16-03-52/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ac5b7b1bc9df640a8af586265eacdbee8c5bbbf7
--- /dev/null
+++ b/outputs/2025-12-19/16-03-52/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219_turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/16-03-52/.hydra/hydra.yaml b/outputs/2025-12-19/16-03-52/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e4a5598b624163e3dd25f7586846a8050410f25e
--- /dev/null
+++ b/outputs/2025-12-19/16-03-52/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219_turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219_turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/16-03-52
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/16-03-52/.hydra/overrides.yaml b/outputs/2025-12-19/16-03-52/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d6691a129d5b3f7eaa3297f2ee5b453aa4232d90
--- /dev/null
+++ b/outputs/2025-12-19/16-03-52/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219_turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/16-03-52/train.log b/outputs/2025-12-19/16-03-52/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/16-46-23/.hydra/config.yaml b/outputs/2025-12-19/16-46-23/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3d9d8024064b0c1111d06b72fdbfe1034e699149
--- /dev/null
+++ b/outputs/2025-12-19/16-46-23/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/16-46-23/.hydra/hydra.yaml b/outputs/2025-12-19/16-46-23/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..06561395c8d5025157083f999f47b3cb5e55630b
--- /dev/null
+++ b/outputs/2025-12-19/16-46-23/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/16-46-23
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/16-46-23/.hydra/overrides.yaml b/outputs/2025-12-19/16-46-23/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..207405e27eca696ef1eb29501eeee9d723326bf1
--- /dev/null
+++ b/outputs/2025-12-19/16-46-23/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube2_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- trainer.experiment_name=rubikscube-2-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/16-46-23/train.log b/outputs/2025-12-19/16-46-23/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/17-34-17/.hydra/config.yaml b/outputs/2025-12-19/17-34-17/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fbfd2fc28c807d6d53f97e714bee1173f07cc4e6
--- /dev/null
+++ b/outputs/2025-12-19/17-34-17/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219-turn5
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/17-34-17/.hydra/hydra.yaml b/outputs/2025-12-19/17-34-17/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e8cc0bfae39ac3ff2346f1e032b81aba19100a4d
--- /dev/null
+++ b/outputs/2025-12-19/17-34-17/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/17-34-17
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/17-34-17/.hydra/overrides.yaml b/outputs/2025-12-19/17-34-17/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c05022799344824b9ea185882e3be8dd53afc7ac
--- /dev/null
+++ b/outputs/2025-12-19/17-34-17/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/17-34-17/train.log b/outputs/2025-12-19/17-34-17/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/20-59-30/.hydra/config.yaml b/outputs/2025-12-19/20-59-30/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..483d702f5de6c317696a764087902c71edd0cede
--- /dev/null
+++ b/outputs/2025-12-19/20-59-30/.hydra/config.yaml
@@ -0,0 +1,358 @@
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+enable_response_mask: true
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_6000
+lora:
+  rank: 0
+  alpha: 64
+  target_modules: all-linear
+actor_rollout_ref:
+  model:
+    path: ${model_path}
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  rollout:
+    name: vllm
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    tensor_model_parallel_size: 1
+    dtype: bfloat16
+    max_model_len: 7200
+    prompt_length: 1
+    response_length: 400
+    gpu_memory_utilization: 0.9
+    max_num_batched_tokens: 8192
+    enforce_eager: true
+    free_cache_engine: true
+    enable_chunked_prefill: false
+    disable_log_stats: false
+    do_sample: true
+    val_kwargs:
+      do_sample: true
+      temperature: 0.5
+      top_p: 1.0
+      top_k: -1
+      logprobs: 20
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 128
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+output:
+  dir: results/eval
+  filename: val_rollouts.pkl
+  append_timestamp: true
+  keep_batch_keys: null
+  keep_non_tensor_keys: null
+  keep_meta_info: true
diff --git a/outputs/2025-12-19/20-59-30/.hydra/hydra.yaml b/outputs/2025-12-19/20-59-30/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cefcc97f21b3c662f95e50de71be196603952f14
--- /dev/null
+++ b/outputs/2025-12-19/20-59-30/.hydra/hydra.yaml
@@ -0,0 +1,157 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: agent_proxy
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: eval.yaml
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/20-59-30
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/20-59-30/.hydra/overrides.yaml b/outputs/2025-12-19/20-59-30/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-19/20-59-30/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-19/20-59-30/agent_proxy.log b/outputs/2025-12-19/20-59-30/agent_proxy.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/21-00-09/.hydra/config.yaml b/outputs/2025-12-19/21-00-09/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..61a4bda688ff45081b1fc70f45b6479d5270b12f
--- /dev/null
+++ b/outputs/2025-12-19/21-00-09/.hydra/config.yaml
@@ -0,0 +1,358 @@
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+enable_response_mask: true
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5_6000
+lora:
+  rank: 0
+  alpha: 64
+  target_modules: all-linear
+actor_rollout_ref:
+  model:
+    path: ${model_path}
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  rollout:
+    name: vllm
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    tensor_model_parallel_size: 1
+    dtype: bfloat16
+    max_model_len: 7200
+    prompt_length: 1
+    response_length: 400
+    gpu_memory_utilization: 0.9
+    max_num_batched_tokens: 8192
+    enforce_eager: true
+    free_cache_engine: true
+    enable_chunked_prefill: false
+    disable_log_stats: false
+    do_sample: true
+    val_kwargs:
+      do_sample: true
+      temperature: 0.5
+      top_p: 1.0
+      top_k: -1
+      logprobs: 20
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 128
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+output:
+  dir: results/eval
+  filename: val_rollouts.pkl
+  append_timestamp: true
+  keep_batch_keys: null
+  keep_non_tensor_keys: null
+  keep_meta_info: true
diff --git a/outputs/2025-12-19/21-00-09/.hydra/hydra.yaml b/outputs/2025-12-19/21-00-09/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6e8bf244236065c2388efebfd5c33f87ca307487
--- /dev/null
+++ b/outputs/2025-12-19/21-00-09/.hydra/hydra.yaml
@@ -0,0 +1,157 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: agent_proxy
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: eval.yaml
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/21-00-09
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/21-00-09/.hydra/overrides.yaml b/outputs/2025-12-19/21-00-09/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-19/21-00-09/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-19/21-00-09/agent_proxy.log b/outputs/2025-12-19/21-00-09/agent_proxy.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/21-01-06/.hydra/config.yaml b/outputs/2025-12-19/21-01-06/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..095508257ebe98c8faefceddff225f7b92cea352
--- /dev/null
+++ b/outputs/2025-12-19/21-01-06/.hydra/config.yaml
@@ -0,0 +1,358 @@
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+enable_response_mask: true
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6
+seed:
+  train: 10000
+  val: 123
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5_6000
+lora:
+  rank: 0
+  alpha: 64
+  target_modules: all-linear
+actor_rollout_ref:
+  model:
+    path: ${model_path}
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  rollout:
+    name: vllm
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    tensor_model_parallel_size: 1
+    dtype: bfloat16
+    max_model_len: 7200
+    prompt_length: 1
+    response_length: 400
+    gpu_memory_utilization: 0.9
+    max_num_batched_tokens: 8192
+    enforce_eager: true
+    free_cache_engine: true
+    enable_chunked_prefill: false
+    disable_log_stats: false
+    do_sample: true
+    val_kwargs:
+      do_sample: true
+      temperature: 0.5
+      top_p: 1.0
+      top_k: -1
+      logprobs: 20
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 128
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+output:
+  dir: results/eval
+  filename: val_rollouts.pkl
+  append_timestamp: true
+  keep_batch_keys: null
+  keep_non_tensor_keys: null
+  keep_meta_info: true
diff --git a/outputs/2025-12-19/21-01-06/.hydra/hydra.yaml b/outputs/2025-12-19/21-01-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fbfea3104d36c32a51efc05077a3010114965546
--- /dev/null
+++ b/outputs/2025-12-19/21-01-06/.hydra/hydra.yaml
@@ -0,0 +1,157 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: agent_proxy
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: eval.yaml
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/21-01-06
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/21-01-06/.hydra/overrides.yaml b/outputs/2025-12-19/21-01-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-19/21-01-06/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-19/21-01-06/agent_proxy.log b/outputs/2025-12-19/21-01-06/agent_proxy.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/21-02-53/.hydra/config.yaml b/outputs/2025-12-19/21-02-53/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..61a4bda688ff45081b1fc70f45b6479d5270b12f
--- /dev/null
+++ b/outputs/2025-12-19/21-02-53/.hydra/config.yaml
@@ -0,0 +1,358 @@
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+enable_response_mask: true
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5_6000
+lora:
+  rank: 0
+  alpha: 64
+  target_modules: all-linear
+actor_rollout_ref:
+  model:
+    path: ${model_path}
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  rollout:
+    name: vllm
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    tensor_model_parallel_size: 1
+    dtype: bfloat16
+    max_model_len: 7200
+    prompt_length: 1
+    response_length: 400
+    gpu_memory_utilization: 0.9
+    max_num_batched_tokens: 8192
+    enforce_eager: true
+    free_cache_engine: true
+    enable_chunked_prefill: false
+    disable_log_stats: false
+    do_sample: true
+    val_kwargs:
+      do_sample: true
+      temperature: 0.5
+      top_p: 1.0
+      top_k: -1
+      logprobs: 20
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 128
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+output:
+  dir: results/eval
+  filename: val_rollouts.pkl
+  append_timestamp: true
+  keep_batch_keys: null
+  keep_non_tensor_keys: null
+  keep_meta_info: true
diff --git a/outputs/2025-12-19/21-02-53/.hydra/hydra.yaml b/outputs/2025-12-19/21-02-53/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1b604a296aa23e8b5dfa1786e0a4db90ec085688
--- /dev/null
+++ b/outputs/2025-12-19/21-02-53/.hydra/hydra.yaml
@@ -0,0 +1,157 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: agent_proxy
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: eval.yaml
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/21-02-53
+    choices:
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/21-02-53/.hydra/overrides.yaml b/outputs/2025-12-19/21-02-53/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-19/21-02-53/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-19/21-02-53/agent_proxy.log b/outputs/2025-12-19/21-02-53/agent_proxy.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/21-14-45/.hydra/config.yaml b/outputs/2025-12-19/21-14-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..07a72caca679064359fc7205e5aceb5d10d3ff9b
--- /dev/null
+++ b/outputs/2025-12-19/21-14-45/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219-turn5_6000
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5_6000
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/21-14-45/.hydra/hydra.yaml b/outputs/2025-12-19/21-14-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f9a47fe8c168a2f94c4d00e229a2b3ca07bd9e40
--- /dev/null
+++ b/outputs/2025-12-19/21-14-45/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5_6000
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5_6000
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5_6000,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5_6000,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/21-14-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/21-14-45/.hydra/overrides.yaml b/outputs/2025-12-19/21-14-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e54b684b1757ffc2eb8f25bc9ceadf124041ea08
--- /dev/null
+++ b/outputs/2025-12-19/21-14-45/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_rubikscube3_1219_turn5_6000
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5_6000
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/21-14-45/train.log b/outputs/2025-12-19/21-14-45/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/21-52-02/.hydra/config.yaml b/outputs/2025-12-19/21-52-02/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d1090d449527f6518163dd4199fc689232302dc2
--- /dev/null
+++ b/outputs/2025-12-19/21-52-02/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219_turn5_6000
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5_6000
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/21-52-02/.hydra/hydra.yaml b/outputs/2025-12-19/21-52-02/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..070dc529e02d21b233e1d1e62322a3da5401379d
--- /dev/null
+++ b/outputs/2025-12-19/21-52-02/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5_6000
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219_turn5_6000
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5_6000,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219_turn5_6000,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/21-52-02
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/21-52-02/.hydra/overrides.yaml b/outputs/2025-12-19/21-52-02/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7bdcf6e6ff9b2a109869492b3cbdbaf61ffa59b3
--- /dev/null
+++ b/outputs/2025-12-19/21-52-02/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_1.5B_it_rubikscube3_1219_turn5_6000
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219_turn5_6000
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/21-52-02/train.log b/outputs/2025-12-19/21-52-02/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-19/22-33-52/.hydra/config.yaml b/outputs/2025-12-19/22-33-52/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..7026aa29cfbff1f8e3361b18bbdf56a632b35d9d
--- /dev/null
+++ b/outputs/2025-12-19/22-33-52/.hydra/config.yaml
@@ -0,0 +1,857 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-frommlp-1219-turn5_6000
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5_6000
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-19/22-33-52/.hydra/hydra.yaml b/outputs/2025-12-19/22-33-52/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..330ea1a269878368bf795e0611a83f5919787fc1
--- /dev/null
+++ b/outputs/2025-12-19/22-33-52/.hydra/hydra.yaml
@@ -0,0 +1,180 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5_6000
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5_6000
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5_6000,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5_6000,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-19/22-33-52
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-19/22-33-52/.hydra/overrides.yaml b/outputs/2025-12-19/22-33-52/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2eee8c887d55a14d72912968711e42569a5f7fb0
--- /dev/null
+++ b/outputs/2025-12-19/22-33-52/.hydra/overrides.yaml
@@ -0,0 +1,11 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_3B_it_rubikscube3_1219_turn5_6000
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- trainer.experiment_name=rubikscube-3-starpos-frommlp-1219-turn5_6000
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-19/22-33-52/train.log b/outputs/2025-12-19/22-33-52/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/01-56-20/.hydra/config.yaml b/outputs/2025-12-20/01-56-20/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..55f8801f6ac366b6ba1003901ec337333adc39c1
--- /dev/null
+++ b/outputs/2025-12-20/01-56-20/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/01-56-20/.hydra/hydra.yaml b/outputs/2025-12-20/01-56-20/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b8d4947d267b4150d164e149bcbe3ef8fd0b517e
--- /dev/null
+++ b/outputs/2025-12-20/01-56-20/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-1-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-1-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/01-56-20
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/01-56-20/.hydra/overrides.yaml b/outputs/2025-12-20/01-56-20/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f38df3b0fa78f53cec079e82b57d76dd39512fb8
--- /dev/null
+++ b/outputs/2025-12-20/01-56-20/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-1-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/01-56-20/train.log b/outputs/2025-12-20/01-56-20/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/01-57-47/.hydra/config.yaml b/outputs/2025-12-20/01-57-47/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..55f8801f6ac366b6ba1003901ec337333adc39c1
--- /dev/null
+++ b/outputs/2025-12-20/01-57-47/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-1-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/01-57-47/.hydra/hydra.yaml b/outputs/2025-12-20/01-57-47/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4b690eff67dd52652c40ef7ad130f2677753d5b3
--- /dev/null
+++ b/outputs/2025-12-20/01-57-47/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-1-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-1-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/01-57-47
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/01-57-47/.hydra/overrides.yaml b/outputs/2025-12-20/01-57-47/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f38df3b0fa78f53cec079e82b57d76dd39512fb8
--- /dev/null
+++ b/outputs/2025-12-20/01-57-47/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-1-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/01-57-47/train.log b/outputs/2025-12-20/01-57-47/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/01-57-56/.hydra/config.yaml b/outputs/2025-12-20/01-57-56/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f993bb48e1f56141175859e85386657f3bbaa477
--- /dev/null
+++ b/outputs/2025-12-20/01-57-56/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/01-57-56/.hydra/hydra.yaml b/outputs/2025-12-20/01-57-56/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..78a696a9b241c3b4653f29f4bf4199a0db9f30cf
--- /dev/null
+++ b/outputs/2025-12-20/01-57-56/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-2-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-2-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/01-57-56
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/01-57-56/.hydra/overrides.yaml b/outputs/2025-12-20/01-57-56/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..62a0c4603e8560a598d9f1ee2fbe7b9916b447d8
--- /dev/null
+++ b/outputs/2025-12-20/01-57-56/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-2-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/01-57-56/train.log b/outputs/2025-12-20/01-57-56/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/01-58-23/.hydra/config.yaml b/outputs/2025-12-20/01-58-23/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..310dfca6a87206cc9e699242310d2ca54ddf0aeb
--- /dev/null
+++ b/outputs/2025-12-20/01-58-23/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: bandit-starpo-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/01-58-23/.hydra/hydra.yaml b/outputs/2025-12-20/01-58-23/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4a532830411c8358782d3527f97285760f042b05
--- /dev/null
+++ b/outputs/2025-12-20/01-58-23/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - trainer.experiment_name=bandit-starpo-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=bandit-starpo-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/01-58-23
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/01-58-23/.hydra/overrides.yaml b/outputs/2025-12-20/01-58-23/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..09291c09ffad6ef057b0fa88c03f7c24211a56cd
--- /dev/null
+++ b/outputs/2025-12-20/01-58-23/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- trainer.experiment_name=bandit-starpo-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/01-58-23/train.log b/outputs/2025-12-20/01-58-23/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/02-38-39/.hydra/config.yaml b/outputs/2025-12-20/02-38-39/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..71d86cbd41d02c121ef4824e0de8a0cbccdc0426
--- /dev/null
+++ b/outputs/2025-12-20/02-38-39/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-starpo-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/02-38-39/.hydra/hydra.yaml b/outputs/2025-12-20/02-38-39/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..14f4f1067b62dc4bd9870f35a61d3a5281888274
--- /dev/null
+++ b/outputs/2025-12-20/02-38-39/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - trainer.experiment_name=bandit-starpo-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=bandit-starpo-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/02-38-39
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/02-38-39/.hydra/overrides.yaml b/outputs/2025-12-20/02-38-39/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9642a56b9ca78ac7376c86f1f693b5d3469b7342
--- /dev/null
+++ b/outputs/2025-12-20/02-38-39/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- trainer.experiment_name=bandit-starpo-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/02-38-39/train.log b/outputs/2025-12-20/02-38-39/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/03-13-25/.hydra/config.yaml b/outputs/2025-12-20/03-13-25/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b50069feacba2ac39adb9f17775f1f25270738d2
--- /dev/null
+++ b/outputs/2025-12-20/03-13-25/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: bandit-starpo-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/03-13-25/.hydra/hydra.yaml b/outputs/2025-12-20/03-13-25/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..78b8a639937f1944d3178a4e83e9b1fa70e8aa24
--- /dev/null
+++ b/outputs/2025-12-20/03-13-25/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - trainer.experiment_name=bandit-starpo-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=bandit-starpo-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/03-13-25
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/03-13-25/.hydra/overrides.yaml b/outputs/2025-12-20/03-13-25/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..225b58098766bb0205a3384b8179bcc52e1284f9
--- /dev/null
+++ b/outputs/2025-12-20/03-13-25/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- trainer.experiment_name=bandit-starpo-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/03-13-25/train.log b/outputs/2025-12-20/03-13-25/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/04-39-39/.hydra/config.yaml b/outputs/2025-12-20/04-39-39/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d927741cdf057453a895430da338e584cc5ab027
--- /dev/null
+++ b/outputs/2025-12-20/04-39-39/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/04-39-39/.hydra/hydra.yaml b/outputs/2025-12-20/04-39-39/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3bf904f37f739a2df9ffa823fa57048a1394f901
--- /dev/null
+++ b/outputs/2025-12-20/04-39-39/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-2-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-2-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/04-39-39
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/04-39-39/.hydra/overrides.yaml b/outputs/2025-12-20/04-39-39/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8980cab8b504cb7f5d3e9aebf0a48ffaaab8c9c5
--- /dev/null
+++ b/outputs/2025-12-20/04-39-39/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-2-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/04-39-39/train.log b/outputs/2025-12-20/04-39-39/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/07-48-38/.hydra/config.yaml b/outputs/2025-12-20/07-48-38/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6872b6e086ab3cdbfd778a577d95e2e2ffc703e2
--- /dev/null
+++ b/outputs/2025-12-20/07-48-38/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-1-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/07-48-38/.hydra/hydra.yaml b/outputs/2025-12-20/07-48-38/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..324240de74645d3201fda222d9958393edc68545
--- /dev/null
+++ b/outputs/2025-12-20/07-48-38/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-1-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-1-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/07-48-38
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/07-48-38/.hydra/overrides.yaml b/outputs/2025-12-20/07-48-38/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d77db62e07c2ec197fcd6b77c2fceab26dfc9f71
--- /dev/null
+++ b/outputs/2025-12-20/07-48-38/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-1-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/07-48-38/train.log b/outputs/2025-12-20/07-48-38/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/08-16-36/.hydra/config.yaml b/outputs/2025-12-20/08-16-36/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a1f47a80c21fc13e904b37461d1b67c50645e763
--- /dev/null
+++ b/outputs/2025-12-20/08-16-36/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-3-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/08-16-36/.hydra/hydra.yaml b/outputs/2025-12-20/08-16-36/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e31180352f355e374b74b377e06d4cfb4ca9679b
--- /dev/null
+++ b/outputs/2025-12-20/08-16-36/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-3-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-3-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/08-16-36
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/08-16-36/.hydra/overrides.yaml b/outputs/2025-12-20/08-16-36/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3ecad84cf48c07af09cc3cadb3a4ad80169c5093
--- /dev/null
+++ b/outputs/2025-12-20/08-16-36/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-3-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/08-16-36/train.log b/outputs/2025-12-20/08-16-36/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/08-20-39/.hydra/config.yaml b/outputs/2025-12-20/08-20-39/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d927741cdf057453a895430da338e584cc5ab027
--- /dev/null
+++ b/outputs/2025-12-20/08-20-39/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/08-20-39/.hydra/hydra.yaml b/outputs/2025-12-20/08-20-39/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b8b50df2f24114e4e313d3470290becc032e7a96
--- /dev/null
+++ b/outputs/2025-12-20/08-20-39/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-2-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-2-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/08-20-39
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/08-20-39/.hydra/overrides.yaml b/outputs/2025-12-20/08-20-39/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8980cab8b504cb7f5d3e9aebf0a48ffaaab8c9c5
--- /dev/null
+++ b/outputs/2025-12-20/08-20-39/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-2-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/08-20-39/train.log b/outputs/2025-12-20/08-20-39/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/08-35-19/.hydra/config.yaml b/outputs/2025-12-20/08-35-19/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5742b88f602a8c2faccc43fa1258afcadc733fa1
--- /dev/null
+++ b/outputs/2025-12-20/08-35-19/.hydra/config.yaml
@@ -0,0 +1,862 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 14400
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: _2048_starpo_stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid.\nMerge equal tiles
+      by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where (0,0) is
+      top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move), a
+      small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>'
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/08-35-19/.hydra/hydra.yaml b/outputs/2025-12-20/08-35-19/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..caf3340478237f0e90531169fdc29a144bfc0795
--- /dev/null
+++ b/outputs/2025-12-20/08-35-19/.hydra/hydra.yaml
@@ -0,0 +1,208 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4
+      grid.\nMerge equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed
+      grid where (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid
+      (no tiles move), a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+    - actor_rollout_ref.rollout.max_model_len=14400
+    - agent_proxy.max_turn=700
+    - trainer.experiment_name=_2048_starpo_stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.max_model_len=14400,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,agent_proxy.max_turn=700,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.game_2048.env_instruction="You
+      are playing the 2048 game on a 4x4 grid.\nMerge equal tiles by sliding Up, Right,
+      Down, or Left.\nUse a zero-indexed grid where (0,0) is top-left and (3,3) is
+      bottom-right.\nIf a move is invalid (no tiles move), a small penalty is applied.\n\nExample
+      answer format:\n<think>\n<observation>\n[Current Grid Symbol Representation]\nTile
+      values and positions: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nI
+      want to merge the two 2s in the top row. Sliding Right will merge them at (0,3).\n<prediction>\n[Predicted
+      Grid Symbol Representation after move]\nPredicted changes: The 2 at (0,0) and
+      2 at (0,1) merge into a 4 at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>",model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=_2048_starpo_stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/08-35-19
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/08-35-19/.hydra/overrides.yaml b/outputs/2025-12-20/08-35-19/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d5aba2427f2592c388a73c15f3046b3471632cb4
--- /dev/null
+++ b/outputs/2025-12-20/08-35-19/.hydra/overrides.yaml
@@ -0,0 +1,26 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4 grid.\nMerge
+  equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where
+  (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move),
+  a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+  Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+  4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+  Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+  after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4 at
+  (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA sample
+  full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n. . . .\n.
+  . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+  moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge to
+  the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile 4 created
+  at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+- actor_rollout_ref.rollout.max_model_len=14400
+- agent_proxy.max_turn=700
+- trainer.experiment_name=_2048_starpo_stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/08-35-19/train.log b/outputs/2025-12-20/08-35-19/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/08-39-07/.hydra/config.yaml b/outputs/2025-12-20/08-39-07/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3f51c07b81fc919583900de52c1e04af320b8646
--- /dev/null
+++ b/outputs/2025-12-20/08-39-07/.hydra/config.yaml
@@ -0,0 +1,862 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 14400
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: _2048_starpo_stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid.\nMerge equal tiles
+      by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where (0,0) is
+      top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move), a
+      small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>'
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/08-39-07/.hydra/hydra.yaml b/outputs/2025-12-20/08-39-07/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d92206de800dc24f77cfc47b4c49c921c9484260
--- /dev/null
+++ b/outputs/2025-12-20/08-39-07/.hydra/hydra.yaml
@@ -0,0 +1,209 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-0.5B-Instruct
+    - 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4
+      grid.\nMerge equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed
+      grid where (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid
+      (no tiles move), a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+    - actor_rollout_ref.rollout.max_model_len=14400
+    - agent_proxy.max_turn=700
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.5
+    - trainer.experiment_name=_2048_starpo_stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.gpu_memory_utilization=0.5,actor_rollout_ref.rollout.max_model_len=14400,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,agent_proxy.max_turn=700,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.game_2048.env_instruction="You
+      are playing the 2048 game on a 4x4 grid.\nMerge equal tiles by sliding Up, Right,
+      Down, or Left.\nUse a zero-indexed grid where (0,0) is top-left and (3,3) is
+      bottom-right.\nIf a move is invalid (no tiles move), a small penalty is applied.\n\nExample
+      answer format:\n<think>\n<observation>\n[Current Grid Symbol Representation]\nTile
+      values and positions: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nI
+      want to merge the two 2s in the top row. Sliding Right will merge them at (0,3).\n<prediction>\n[Predicted
+      Grid Symbol Representation after move]\nPredicted changes: The 2 at (0,0) and
+      2 at (0,1) merge into a 4 at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>",model_path=Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=_2048_starpo_stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/08-39-07
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/08-39-07/.hydra/overrides.yaml b/outputs/2025-12-20/08-39-07/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f1ad4a11cdae19ee8f9eaddeb68a3c6aba25442a
--- /dev/null
+++ b/outputs/2025-12-20/08-39-07/.hydra/overrides.yaml
@@ -0,0 +1,27 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-0.5B-Instruct
+- 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4 grid.\nMerge
+  equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where
+  (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move),
+  a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+  Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+  4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+  Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+  after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4 at
+  (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA sample
+  full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n. . . .\n.
+  . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+  moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge to
+  the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile 4 created
+  at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+- actor_rollout_ref.rollout.max_model_len=14400
+- agent_proxy.max_turn=700
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.5
+- trainer.experiment_name=_2048_starpo_stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/08-39-07/train.log b/outputs/2025-12-20/08-39-07/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/08-54-12/.hydra/config.yaml b/outputs/2025-12-20/08-54-12/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..23c6cc5e3bc837f913ca83c137406b1d3bd32692
--- /dev/null
+++ b/outputs/2025-12-20/08-54-12/.hydra/config.yaml
@@ -0,0 +1,862 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 14400
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: _2048_starpo_stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid.\nMerge equal tiles
+      by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where (0,0) is
+      top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move), a
+      small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>'
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/08-54-12/.hydra/hydra.yaml b/outputs/2025-12-20/08-54-12/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..024460015917034a9b4d580c9d71f443b4e9cfee
--- /dev/null
+++ b/outputs/2025-12-20/08-54-12/.hydra/hydra.yaml
@@ -0,0 +1,209 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+    - 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4
+      grid.\nMerge equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed
+      grid where (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid
+      (no tiles move), a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+    - actor_rollout_ref.rollout.max_model_len=14400
+    - agent_proxy.max_turn=700
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.5
+    - trainer.experiment_name=_2048_starpo_stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.gpu_memory_utilization=0.5,actor_rollout_ref.rollout.max_model_len=14400,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,agent_proxy.max_turn=700,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.game_2048.env_instruction="You
+      are playing the 2048 game on a 4x4 grid.\nMerge equal tiles by sliding Up, Right,
+      Down, or Left.\nUse a zero-indexed grid where (0,0) is top-left and (3,3) is
+      bottom-right.\nIf a move is invalid (no tiles move), a small penalty is applied.\n\nExample
+      answer format:\n<think>\n<observation>\n[Current Grid Symbol Representation]\nTile
+      values and positions: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nI
+      want to merge the two 2s in the top row. Sliding Right will merge them at (0,3).\n<prediction>\n[Predicted
+      Grid Symbol Representation after move]\nPredicted changes: The 2 at (0,0) and
+      2 at (0,1) merge into a 4 at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>",model_path=/mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=_2048_starpo_stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/08-54-12
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/08-54-12/.hydra/overrides.yaml b/outputs/2025-12-20/08-54-12/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f2c159515cd0e60b2050f9209fe97293334c9377
--- /dev/null
+++ b/outputs/2025-12-20/08-54-12/.hydra/overrides.yaml
@@ -0,0 +1,27 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+- 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4 grid.\nMerge
+  equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where
+  (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move),
+  a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+  Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+  4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+  Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+  after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4 at
+  (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA sample
+  full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n. . . .\n.
+  . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+  moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge to
+  the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile 4 created
+  at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+- actor_rollout_ref.rollout.max_model_len=14400
+- agent_proxy.max_turn=700
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.5
+- trainer.experiment_name=_2048_starpo_stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/08-54-12/train.log b/outputs/2025-12-20/08-54-12/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/10-20-24/.hydra/config.yaml b/outputs/2025-12-20/10-20-24/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e5aa946d850c8c4bd7981bc103a5a0b100e46933
--- /dev/null
+++ b/outputs/2025-12-20/10-20-24/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-1-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/10-20-24/.hydra/hydra.yaml b/outputs/2025-12-20/10-20-24/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..11c5ab8acd93e9884bd27097ab6d9f1acb95aab1
--- /dev/null
+++ b/outputs/2025-12-20/10-20-24/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=1
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-1-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=1,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-1-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/10-20-24
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/10-20-24/.hydra/overrides.yaml b/outputs/2025-12-20/10-20-24/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..05efe137bd1cb67f922bd15e12da9dbacf0f17d3
--- /dev/null
+++ b/outputs/2025-12-20/10-20-24/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=1
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-1-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/10-20-24/train.log b/outputs/2025-12-20/10-20-24/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/11-10-46/.hydra/config.yaml b/outputs/2025-12-20/11-10-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bdb9f2b20e9380bdeef598e88a588c92aaa468e5
--- /dev/null
+++ b/outputs/2025-12-20/11-10-46/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-2-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/11-10-46/.hydra/hydra.yaml b/outputs/2025-12-20/11-10-46/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0023de2b1a8b2d8106af35610e740cf150a0c32a
--- /dev/null
+++ b/outputs/2025-12-20/11-10-46/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-2-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-2-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/11-10-46
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/11-10-46/.hydra/overrides.yaml b/outputs/2025-12-20/11-10-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0859bec054ed058baf380189f0b6916df951c2a8
--- /dev/null
+++ b/outputs/2025-12-20/11-10-46/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-2-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/11-10-46/train.log b/outputs/2025-12-20/11-10-46/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/11-23-55/.hydra/config.yaml b/outputs/2025-12-20/11-23-55/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ad7919730f08681226fe88ac1d503ac0acab2b84
--- /dev/null
+++ b/outputs/2025-12-20/11-23-55/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-3-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-1.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/11-23-55/.hydra/hydra.yaml b/outputs/2025-12-20/11-23-55/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3358027672afea86cc5a96df25fc92ddb0f206d0
--- /dev/null
+++ b/outputs/2025-12-20/11-23-55/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-1.5B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-3-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-1.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-3-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/11-23-55
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/11-23-55/.hydra/overrides.yaml b/outputs/2025-12-20/11-23-55/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2ecdb060cfd04bf1f1f66538a3bc5644e6e22082
--- /dev/null
+++ b/outputs/2025-12-20/11-23-55/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-1.5B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-3-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/11-23-55/train.log b/outputs/2025-12-20/11-23-55/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/14-11-39/.hydra/config.yaml b/outputs/2025-12-20/14-11-39/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..adf2d48b0d2b9693936bb26b5df1ad46d755505d
--- /dev/null
+++ b/outputs/2025-12-20/14-11-39/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: rubikscube-2-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-0.5B-text_with_coordinates-qwen/global_step_1335
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/14-11-39/.hydra/hydra.yaml b/outputs/2025-12-20/14-11-39/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..56e5fa1f405200b8200ce5852befacc2b83f7667
--- /dev/null
+++ b/outputs/2025-12-20/14-11-39/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-0.5B-text_with_coordinates-qwen/global_step_1335
+    - custom_envs.rubikscube.env_config.scramble_depth=2
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-2-starpos-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=2,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-0.5B-text_with_coordinates-qwen/global_step_1335,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-2-starpos-SPA,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/14-11-39
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/14-11-39/.hydra/overrides.yaml b/outputs/2025-12-20/14-11-39/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4158031feed84079c08acbe4f1262c2428199c81
--- /dev/null
+++ b/outputs/2025-12-20/14-11-39/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-0.5B-text_with_coordinates-qwen/global_step_1335
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-2-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/14-11-39/train.log b/outputs/2025-12-20/14-11-39/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/14-23-16/.hydra/config.yaml b/outputs/2025-12-20/14-23-16/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0aaf3465c946e52a6386d0db026532014b7258c1
--- /dev/null
+++ b/outputs/2025-12-20/14-23-16/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: Qwen/Qwen2.5-3B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/14-23-16/.hydra/hydra.yaml b/outputs/2025-12-20/14-23-16/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0b72dbdafdd391f63912b94f65991bca50984050
--- /dev/null
+++ b/outputs/2025-12-20/14-23-16/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=Qwen/Qwen2.5-3B-Instruct
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-3-starpos-stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=Qwen/Qwen2.5-3B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-3-starpos-stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/14-23-16
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/14-23-16/.hydra/overrides.yaml b/outputs/2025-12-20/14-23-16/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b48b63a208ec9f7a35d665a01da2be4fd065b976
--- /dev/null
+++ b/outputs/2025-12-20/14-23-16/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=Qwen/Qwen2.5-3B-Instruct
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-3-starpos-stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/14-23-16/train.log b/outputs/2025-12-20/14-23-16/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/14-53-36/.hydra/config.yaml b/outputs/2025-12-20/14-53-36/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..def008b81cfc4ba592cd13d19eb3f5138770d184
--- /dev/null
+++ b/outputs/2025-12-20/14-53-36/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: bandit-starpo-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/14-53-36/.hydra/hydra.yaml b/outputs/2025-12-20/14-53-36/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ed4e8a8537e4f68e0f829c86c7283141855af06d
--- /dev/null
+++ b/outputs/2025-12-20/14-53-36/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen
+    - trainer.experiment_name=bandit-starpo-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=bandit-starpo-SPA,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/14-53-36
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/14-53-36/.hydra/overrides.yaml b/outputs/2025-12-20/14-53-36/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..272be056263d7dc016f31d19083f4345fcc13168
--- /dev/null
+++ b/outputs/2025-12-20/14-53-36/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen
+- trainer.experiment_name=bandit-starpo-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/14-53-36/train.log b/outputs/2025-12-20/14-53-36/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/14-55-23/.hydra/config.yaml b/outputs/2025-12-20/14-55-23/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..cbb2ffbf75e46d5e4e78d51970c576a6d6f9bff5
--- /dev/null
+++ b/outputs/2025-12-20/14-55-23/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: bandit-starpo-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen/global_step_320
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/14-55-23/.hydra/hydra.yaml b/outputs/2025-12-20/14-55-23/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b7b9c3e8513c68518064bc38996f53dbb8a8b933
--- /dev/null
+++ b/outputs/2025-12-20/14-55-23/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen/global_step_320
+    - trainer.experiment_name=bandit-starpo-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen/global_step_320,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=bandit-starpo-SPA,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/14-55-23
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/14-55-23/.hydra/overrides.yaml b/outputs/2025-12-20/14-55-23/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a634b855a145443f98aa19cf6c3f2c0f853c0d14
--- /dev/null
+++ b/outputs/2025-12-20/14-55-23/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-0.5B-text_with_coordinates-qwen/global_step_320
+- trainer.experiment_name=bandit-starpo-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/14-55-23/train.log b/outputs/2025-12-20/14-55-23/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/15-18-26/.hydra/config.yaml b/outputs/2025-12-20/15-18-26/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9f32686b000185d0d1484dfb7a6fc156de7f7863
--- /dev/null
+++ b/outputs/2025-12-20/15-18-26/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: bandit-starpo-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-1.5B-text_with_coordinates-qwen/global_step_320
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/15-18-26/.hydra/hydra.yaml b/outputs/2025-12-20/15-18-26/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b0c30aeb7f5c76e91349b6a8d6326a7d97fd31da
--- /dev/null
+++ b/outputs/2025-12-20/15-18-26/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_15b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-1.5B-text_with_coordinates-qwen/global_step_320
+    - trainer.experiment_name=bandit-starpo-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-1.5B-text_with_coordinates-qwen/global_step_320,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=bandit-starpo-SPA,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_15b_it
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/15-18-26
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/15-18-26/.hydra/overrides.yaml b/outputs/2025-12-20/15-18-26/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..24ea2a1313627c0dcec00329ebc589ec9d90bf7b
--- /dev/null
+++ b/outputs/2025-12-20/15-18-26/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-1.5B-text_with_coordinates-qwen/global_step_320
+- trainer.experiment_name=bandit-starpo-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/15-18-26/train.log b/outputs/2025-12-20/15-18-26/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/15-48-20/.hydra/config.yaml b/outputs/2025-12-20/15-48-20/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f2f9bc16186dfa65d043ee471f63ae813963c15
--- /dev/null
+++ b/outputs/2025-12-20/15-48-20/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_15b_it
+  experiment_name: rubikscube-2-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-1.5B-text_with_coordinates-qwen/global_step_1335
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/15-48-20/.hydra/overrides.yaml b/outputs/2025-12-20/15-48-20/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e63ec7505090bc683aed42884791ed78e45c4dd2
--- /dev/null
+++ b/outputs/2025-12-20/15-48-20/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-1.5B-text_with_coordinates-qwen/global_step_1335
+- custom_envs.rubikscube.env_config.scramble_depth=2
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-2-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/15-48-20/train.log b/outputs/2025-12-20/15-48-20/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/15-49-19/.hydra/config.yaml b/outputs/2025-12-20/15-49-19/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dbb98b8cbafb6de3662408d59c4482316f24555d
--- /dev/null
+++ b/outputs/2025-12-20/15-49-19/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: bandit-starpo-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-3B-text_with_coordinates-qwen/global_step_320
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 1
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - Bandit
+      n_groups:
+      - 8
+  val:
+    env_groups: 512
+    group_size: 1
+    env_configs:
+      tags:
+      - Bandit
+      - BanditTest
+      n_groups:
+      - 256
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-20/15-49-19/.hydra/hydra.yaml b/outputs/2025-12-20/15-49-19/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3a393cbf635e7dfda28410fd5bb5a8eca57c1ce2
--- /dev/null
+++ b/outputs/2025-12-20/15-49-19/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-3B-text_with_coordinates-qwen/global_step_320
+    - trainer.experiment_name=bandit-starpo-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-3B-text_with_coordinates-qwen/global_step_320,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=bandit-starpo-SPA,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it
+    id: ???
+    num: ???
+    config_name: _1_bandit
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/15-49-19
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/15-49-19/.hydra/overrides.yaml b/outputs/2025-12-20/15-49-19/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5f58c6405fe464a7b2a0cfca245e781f891119c5
--- /dev/null
+++ b/outputs/2025-12-20/15-49-19/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_1_bandit-3B-text_with_coordinates-qwen/global_step_320
+- trainer.experiment_name=bandit-starpo-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-20/15-49-19/train.log b/outputs/2025-12-20/15-49-19/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/16-22-49/.hydra/config.yaml b/outputs/2025-12-20/16-22-49/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ef206dbafac2fd19f6bb7e1377b598628c95668d
--- /dev/null
+++ b/outputs/2025-12-20/16-22-49/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gpt-4o-mini
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-22-49/.hydra/hydra.yaml b/outputs/2025-12-20/16-22-49/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ea2c3717851bcfb4e4ad0cd17ad5b69b81f28194
--- /dev/null
+++ b/outputs/2025-12-20/16-22-49/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-22-49
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-22-49/.hydra/overrides.yaml b/outputs/2025-12-20/16-22-49/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/16-22-49/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/16-22-49/eval_api.log b/outputs/2025-12-20/16-22-49/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..4bda3ee00c5809659b0d1c6773b45b972696484b
--- /dev/null
+++ b/outputs/2025-12-20/16-22-49/eval_api.log
@@ -0,0 +1,6 @@
+[2025-12-20 16:22:52,444][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:22:52,888][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:22:52,917][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:22:52,992][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:22:53,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:22:53,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-20/16-23-25/.hydra/config.yaml b/outputs/2025-12-20/16-23-25/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9368dee94e7298be88fd2c004651aead7c390114
--- /dev/null
+++ b/outputs/2025-12-20/16-23-25/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-23-25/.hydra/hydra.yaml b/outputs/2025-12-20/16-23-25/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..6fdfcbdf55cc9e014ad3c408619ad520549667fa
--- /dev/null
+++ b/outputs/2025-12-20/16-23-25/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-23-25
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-23-25/.hydra/overrides.yaml b/outputs/2025-12-20/16-23-25/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/16-23-25/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/16-23-25/eval_api.log b/outputs/2025-12-20/16-23-25/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..d620ae1a4e0eb7fbb5469490207845c20283bbca
--- /dev/null
+++ b/outputs/2025-12-20/16-23-25/eval_api.log
@@ -0,0 +1,77 @@
+[2025-12-20 16:23:31,527][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433798 seconds
+[2025-12-20 16:23:31,528][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405480 seconds
+[2025-12-20 16:23:31,528][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402372 seconds
+[2025-12-20 16:23:31,870][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:32,288][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904805 seconds
+[2025-12-20 16:23:32,663][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:32,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:33,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:33,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:35,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:36,935][openai._base_client][INFO] - Retrying request to /chat/completions in 0.829006 seconds
+[2025-12-20 16:23:36,964][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955999 seconds
+[2025-12-20 16:23:39,935][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:40,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:41,513][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:41,699][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:41,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:42,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:43,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:44,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:45,073][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:45,491][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:46,037][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:49,860][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:49,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:50,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:53,665][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:55,642][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:56,270][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:56,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:23:58,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:01,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:01,181][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:05,696][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:06,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:06,989][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:12,124][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397819 seconds
+[2025-12-20 16:24:12,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788044 seconds
+[2025-12-20 16:24:15,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:19,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:24:24,583][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410936 seconds
+[2025-12-20 16:24:30,001][openai._base_client][INFO] - Retrying request to /chat/completions in 0.895680 seconds
+[2025-12-20 16:24:45,918][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406660 seconds
+[2025-12-20 16:24:51,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901645 seconds
+[2025-12-20 16:25:04,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:04,943][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443634 seconds
+[2025-12-20 16:25:09,943][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475226 seconds
+[2025-12-20 16:25:09,946][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446872 seconds
+[2025-12-20 16:25:11,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:11,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:11,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:12,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:12,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:13,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:15,424][openai._base_client][INFO] - Retrying request to /chat/completions in 0.964726 seconds
+[2025-12-20 16:25:16,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:18,335][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:18,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:22,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:22,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:24,303][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:25,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:26,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:27,609][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:27,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:28,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:29,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:29,475][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:29,816][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:35,139][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:41,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:43,953][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:44,365][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:46,643][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:54,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:57,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:25:59,596][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-20/16-26-49/.hydra/config.yaml b/outputs/2025-12-20/16-26-49/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9368dee94e7298be88fd2c004651aead7c390114
--- /dev/null
+++ b/outputs/2025-12-20/16-26-49/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-26-49/.hydra/hydra.yaml b/outputs/2025-12-20/16-26-49/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b890bdcebebe755619a01f9385e6579c0bddf101
--- /dev/null
+++ b/outputs/2025-12-20/16-26-49/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-26-49
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-26-49/.hydra/overrides.yaml b/outputs/2025-12-20/16-26-49/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/16-26-49/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/16-26-49/eval_api.log b/outputs/2025-12-20/16-26-49/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..4406e1658c06054a1a2c3bee40a5421635452c77
--- /dev/null
+++ b/outputs/2025-12-20/16-26-49/eval_api.log
@@ -0,0 +1,36 @@
+[2025-12-20 16:26:50,972][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:50,989][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,006][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,022][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,037][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,053][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,072][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,089][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,104][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,120][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,136][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,153][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,169][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,186][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,201][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,218][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,234][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,316][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,395][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,396][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,399][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:51,415][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:52,073][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:52,081][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:52,091][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:52,186][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:52,189][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:52,247][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:55,762][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486132 seconds
+[2025-12-20 16:26:55,763][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444519 seconds
+[2025-12-20 16:26:55,763][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432230 seconds
+[2025-12-20 16:26:55,764][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404973 seconds
+[2025-12-20 16:26:56,186][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:56,212][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:56,224][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:26:56,266][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
diff --git a/outputs/2025-12-20/16-27-49/.hydra/config.yaml b/outputs/2025-12-20/16-27-49/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..9368dee94e7298be88fd2c004651aead7c390114
--- /dev/null
+++ b/outputs/2025-12-20/16-27-49/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-27-49/.hydra/hydra.yaml b/outputs/2025-12-20/16-27-49/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f1d0a5645a1d6cccbd31da8084b2901003cb816a
--- /dev/null
+++ b/outputs/2025-12-20/16-27-49/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-27-49
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-27-49/.hydra/overrides.yaml b/outputs/2025-12-20/16-27-49/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/16-27-49/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/16-27-49/eval_api.log b/outputs/2025-12-20/16-27-49/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..3245a4a8e14283cf434eb5cb8d48c690afe59420
--- /dev/null
+++ b/outputs/2025-12-20/16-27-49/eval_api.log
@@ -0,0 +1,32 @@
+[2025-12-20 16:27:50,482][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,485][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,487][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,488][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,489][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,490][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,490][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,496][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,497][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,500][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,502][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,504][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,508][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,509][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,514][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,516][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,520][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,526][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,527][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,527][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,529][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,529][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,530][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,533][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,534][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,536][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,537][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,538][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,552][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,560][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,561][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:27:50,568][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
diff --git a/outputs/2025-12-20/16-28-13/.hydra/config.yaml b/outputs/2025-12-20/16-28-13/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..336b71034095ff89a01394677934cc07ace997b2
--- /dev/null
+++ b/outputs/2025-12-20/16-28-13/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 8192
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-28-13/.hydra/hydra.yaml b/outputs/2025-12-20/16-28-13/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2a34cf6af67acde135f32dec89da42a0bd08e9d5
--- /dev/null
+++ b/outputs/2025-12-20/16-28-13/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-28-13
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-28-13/.hydra/overrides.yaml b/outputs/2025-12-20/16-28-13/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/16-28-13/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/16-28-13/eval_api.log b/outputs/2025-12-20/16-28-13/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..87e127326a957b4f819e505ccc22b5438c3af647
--- /dev/null
+++ b/outputs/2025-12-20/16-28-13/eval_api.log
@@ -0,0 +1,123 @@
+[2025-12-20 16:28:15,319][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458719 seconds
+[2025-12-20 16:28:15,322][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386460 seconds
+[2025-12-20 16:28:15,324][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384133 seconds
+[2025-12-20 16:28:15,324][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411100 seconds
+[2025-12-20 16:28:19,753][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:19,969][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410897 seconds
+[2025-12-20 16:28:19,969][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402295 seconds
+[2025-12-20 16:28:19,972][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389866 seconds
+[2025-12-20 16:28:20,711][openai._base_client][INFO] - Retrying request to /chat/completions in 0.935851 seconds
+[2025-12-20 16:28:20,711][openai._base_client][INFO] - Retrying request to /chat/completions in 0.879477 seconds
+[2025-12-20 16:28:20,739][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860098 seconds
+[2025-12-20 16:28:20,780][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816067 seconds
+[2025-12-20 16:28:22,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:23,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:24,711][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:24,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:25,365][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806948 seconds
+[2025-12-20 16:28:25,383][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869180 seconds
+[2025-12-20 16:28:26,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:27,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:28,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:28,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:31,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:32,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:33,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:34,147][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:35,596][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:36,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:38,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:38,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:39,139][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:42,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:44,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:46,299][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:47,981][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:48,232][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:49,055][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:51,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:51,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:52,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:53,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:28:53,639][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:01,509][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:06,737][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468503 seconds
+[2025-12-20 16:29:11,743][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376700 seconds
+[2025-12-20 16:29:14,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:23,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:23,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459666 seconds
+[2025-12-20 16:29:27,824][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:28,950][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492666 seconds
+[2025-12-20 16:29:28,950][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456065 seconds
+[2025-12-20 16:29:28,950][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493738 seconds
+[2025-12-20 16:29:28,951][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403953 seconds
+[2025-12-20 16:29:28,951][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489956 seconds
+[2025-12-20 16:29:28,951][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476153 seconds
+[2025-12-20 16:29:28,952][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408370 seconds
+[2025-12-20 16:29:28,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408851 seconds
+[2025-12-20 16:29:28,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469558 seconds
+[2025-12-20 16:29:28,954][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437537 seconds
+[2025-12-20 16:29:28,954][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446538 seconds
+[2025-12-20 16:29:29,411][openai._base_client][INFO] - Retrying request to /chat/completions in 0.798388 seconds
+[2025-12-20 16:29:34,361][openai._base_client][INFO] - Retrying request to /chat/completions in 0.782871 seconds
+[2025-12-20 16:29:34,362][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908804 seconds
+[2025-12-20 16:29:34,364][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812100 seconds
+[2025-12-20 16:29:34,394][openai._base_client][INFO] - Retrying request to /chat/completions in 0.764445 seconds
+[2025-12-20 16:29:34,402][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844775 seconds
+[2025-12-20 16:29:34,408][openai._base_client][INFO] - Retrying request to /chat/completions in 0.765183 seconds
+[2025-12-20 16:29:34,424][openai._base_client][INFO] - Retrying request to /chat/completions in 0.977025 seconds
+[2025-12-20 16:29:34,430][openai._base_client][INFO] - Retrying request to /chat/completions in 0.771467 seconds
+[2025-12-20 16:29:34,442][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905180 seconds
+[2025-12-20 16:29:34,444][openai._base_client][INFO] - Retrying request to /chat/completions in 0.891011 seconds
+[2025-12-20 16:29:34,446][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815344 seconds
+[2025-12-20 16:29:40,215][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461480 seconds
+[2025-12-20 16:29:45,158][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386073 seconds
+[2025-12-20 16:29:45,163][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418486 seconds
+[2025-12-20 16:29:45,206][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439755 seconds
+[2025-12-20 16:29:45,251][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440361 seconds
+[2025-12-20 16:29:45,267][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475526 seconds
+[2025-12-20 16:29:45,408][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388453 seconds
+[2025-12-20 16:29:45,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496668 seconds
+[2025-12-20 16:29:45,679][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916396 seconds
+[2025-12-20 16:29:47,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:48,919][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:49,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:50,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:50,372][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:50,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.819946 seconds
+[2025-12-20 16:29:50,648][openai._base_client][INFO] - Retrying request to /chat/completions in 0.754453 seconds
+[2025-12-20 16:29:50,799][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788589 seconds
+[2025-12-20 16:29:51,665][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:52,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:54,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:55,394][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:55,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:29:58,584][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:00,108][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:00,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:02,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:06,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:18,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:20,226][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:25,454][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411433 seconds
+[2025-12-20 16:30:29,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:30,473][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409986 seconds
+[2025-12-20 16:30:30,474][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380315 seconds
+[2025-12-20 16:30:30,474][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377826 seconds
+[2025-12-20 16:30:30,475][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435731 seconds
+[2025-12-20 16:30:30,475][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388494 seconds
+[2025-12-20 16:30:30,475][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389992 seconds
+[2025-12-20 16:30:30,475][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421391 seconds
+[2025-12-20 16:30:30,870][openai._base_client][INFO] - Retrying request to /chat/completions in 0.770650 seconds
+[2025-12-20 16:30:35,859][openai._base_client][INFO] - Retrying request to /chat/completions in 0.930265 seconds
+[2025-12-20 16:30:35,865][openai._base_client][INFO] - Retrying request to /chat/completions in 0.808656 seconds
+[2025-12-20 16:30:35,867][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781788 seconds
+[2025-12-20 16:30:35,887][openai._base_client][INFO] - Retrying request to /chat/completions in 0.780551 seconds
+[2025-12-20 16:30:35,899][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833770 seconds
+[2025-12-20 16:30:35,911][openai._base_client][INFO] - Retrying request to /chat/completions in 0.895942 seconds
+[2025-12-20 16:30:37,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:37,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:39,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:30:43,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:31:01,055][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:31:04,504][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-20/16-31-29/.hydra/config.yaml b/outputs/2025-12-20/16-31-29/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a4493e0d1a7ae550a78eecdaed88f301ec048e01
--- /dev/null
+++ b/outputs/2025-12-20/16-31-29/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-31-29/.hydra/hydra.yaml b/outputs/2025-12-20/16-31-29/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ce2b7118c9be90579693adbbe9ef86f30b826f75
--- /dev/null
+++ b/outputs/2025-12-20/16-31-29/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-31-29
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-31-29/.hydra/overrides.yaml b/outputs/2025-12-20/16-31-29/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/16-31-29/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/16-39-55/.hydra/config.yaml b/outputs/2025-12-20/16-39-55/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a0ae7edf93e2eab7410690d45788ce958d67be92
--- /dev/null
+++ b/outputs/2025-12-20/16-39-55/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-39-55/eval_api.log b/outputs/2025-12-20/16-39-55/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..c36e74df5d8d2e5d450a1659e4f86d2d8fe369f1
--- /dev/null
+++ b/outputs/2025-12-20/16-39-55/eval_api.log
@@ -0,0 +1,131 @@
+[2025-12-20 16:39:59,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:02,081][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406743 seconds
+[2025-12-20 16:40:02,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473965 seconds
+[2025-12-20 16:40:02,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429533 seconds
+[2025-12-20 16:40:02,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459811 seconds
+[2025-12-20 16:40:02,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497640 seconds
+[2025-12-20 16:40:02,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378290 seconds
+[2025-12-20 16:40:02,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437600 seconds
+[2025-12-20 16:40:02,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485215 seconds
+[2025-12-20 16:40:02,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464762 seconds
+[2025-12-20 16:40:02,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425322 seconds
+[2025-12-20 16:40:04,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:05,430][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:05,684][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:06,870][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:07,492][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985725 seconds
+[2025-12-20 16:40:07,511][openai._base_client][INFO] - Retrying request to /chat/completions in 0.754209 seconds
+[2025-12-20 16:40:07,513][openai._base_client][INFO] - Retrying request to /chat/completions in 0.918151 seconds
+[2025-12-20 16:40:07,521][openai._base_client][INFO] - Retrying request to /chat/completions in 0.900601 seconds
+[2025-12-20 16:40:07,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899208 seconds
+[2025-12-20 16:40:07,549][openai._base_client][INFO] - Retrying request to /chat/completions in 0.959678 seconds
+[2025-12-20 16:40:07,558][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933702 seconds
+[2025-12-20 16:40:07,570][openai._base_client][INFO] - Retrying request to /chat/completions in 0.914606 seconds
+[2025-12-20 16:40:07,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.843744 seconds
+[2025-12-20 16:40:08,587][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:09,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:10,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:13,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:16,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:20,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:25,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382822 seconds
+[2025-12-20 16:40:30,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421208 seconds
+[2025-12-20 16:40:30,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428129 seconds
+[2025-12-20 16:40:30,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480066 seconds
+[2025-12-20 16:40:30,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454466 seconds
+[2025-12-20 16:40:32,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:35,977][openai._base_client][INFO] - Retrying request to /chat/completions in 0.968027 seconds
+[2025-12-20 16:40:35,999][openai._base_client][INFO] - Retrying request to /chat/completions in 0.965794 seconds
+[2025-12-20 16:40:36,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:38,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:41,209][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:45,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:45,157][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376428 seconds
+[2025-12-20 16:40:45,158][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483614 seconds
+[2025-12-20 16:40:47,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:47,917][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:48,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:49,432][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:49,812][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:50,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471517 seconds
+[2025-12-20 16:40:50,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383143 seconds
+[2025-12-20 16:40:50,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378440 seconds
+[2025-12-20 16:40:50,163][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473985 seconds
+[2025-12-20 16:40:50,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:40:55,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:00,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:00,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:02,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:05,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:06,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:07,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:11,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:11,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:43,303][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:43,669][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451743 seconds
+[2025-12-20 16:41:45,691][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:47,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:48,671][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435395 seconds
+[2025-12-20 16:41:48,672][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405723 seconds
+[2025-12-20 16:41:48,672][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416548 seconds
+[2025-12-20 16:41:48,672][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499661 seconds
+[2025-12-20 16:41:48,673][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465707 seconds
+[2025-12-20 16:41:48,673][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447738 seconds
+[2025-12-20 16:41:48,928][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386726 seconds
+[2025-12-20 16:41:51,337][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:53,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:54,111][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972276 seconds
+[2025-12-20 16:41:54,124][openai._base_client][INFO] - Retrying request to /chat/completions in 0.886122 seconds
+[2025-12-20 16:41:54,317][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807454 seconds
+[2025-12-20 16:41:56,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:57,119][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:41:57,752][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:42:03,798][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:42:05,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:42:06,838][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:42:11,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:42:17,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:42:19,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:42:25,100][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442285 seconds
+[2025-12-20 16:43:13,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:13,782][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387692 seconds
+[2025-12-20 16:43:15,342][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:18,789][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409809 seconds
+[2025-12-20 16:43:18,790][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388791 seconds
+[2025-12-20 16:43:19,174][openai._base_client][INFO] - Retrying request to /chat/completions in 0.786218 seconds
+[2025-12-20 16:43:21,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:22,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:24,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:24,202][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807094 seconds
+[2025-12-20 16:43:24,256][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:24,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:25,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:31,426][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:43:40,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:07,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:19,982][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:33,069][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:38,364][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375692 seconds
+[2025-12-20 16:44:44,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:50,755][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:50,826][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396413 seconds
+[2025-12-20 16:44:53,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:53,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:44:55,829][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386899 seconds
+[2025-12-20 16:44:55,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498736 seconds
+[2025-12-20 16:44:55,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479151 seconds
+[2025-12-20 16:44:56,099][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439822 seconds
+[2025-12-20 16:45:01,337][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999586 seconds
+[2025-12-20 16:45:04,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:05,228][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:06,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:08,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:08,222][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:28,729][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:30,199][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:34,530][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:52,127][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:45:59,283][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:46:04,512][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421631 seconds
+[2025-12-20 16:46:09,941][openai._base_client][INFO] - Retrying request to /chat/completions in 0.804779 seconds
+[2025-12-20 16:46:17,140][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-20/16-49-37/.hydra/config.yaml b/outputs/2025-12-20/16-49-37/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..701e34f20dcc50b6e0f527ca2e2c0ff7229bf9b2
--- /dev/null
+++ b/outputs/2025-12-20/16-49-37/.hydra/config.yaml
@@ -0,0 +1,917 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: TA/openai/gpt-oss-120b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/16-49-37/.hydra/hydra.yaml b/outputs/2025-12-20/16-49-37/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b4eb59cd4af484275e248f29eeba022de90cce69
--- /dev/null
+++ b/outputs/2025-12-20/16-49-37/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-49-37
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-49-37/eval_api.log b/outputs/2025-12-20/16-49-37/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..0ad42cd2e67f32f80387f5d3c6aaf46bc4c612f3
--- /dev/null
+++ b/outputs/2025-12-20/16-49-37/eval_api.log
@@ -0,0 +1,137 @@
+[2025-12-20 16:49:41,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:41,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:41,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:43,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:43,801][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409757 seconds
+[2025-12-20 16:49:43,803][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381520 seconds
+[2025-12-20 16:49:43,804][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430403 seconds
+[2025-12-20 16:49:43,804][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492733 seconds
+[2025-12-20 16:49:43,806][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427296 seconds
+[2025-12-20 16:49:43,807][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379875 seconds
+[2025-12-20 16:49:44,204][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:45,654][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:53,476][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:55,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:55,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:55,500][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:56,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:57,189][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:49:59,775][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:02,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:03,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:18,962][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:19,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444515 seconds
+[2025-12-20 16:50:24,236][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426061 seconds
+[2025-12-20 16:50:24,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401585 seconds
+[2025-12-20 16:50:24,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491475 seconds
+[2025-12-20 16:50:24,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442391 seconds
+[2025-12-20 16:50:24,339][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:24,677][openai._base_client][INFO] - Retrying request to /chat/completions in 0.776224 seconds
+[2025-12-20 16:50:26,170][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:27,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:28,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:29,666][openai._base_client][INFO] - Retrying request to /chat/completions in 0.882733 seconds
+[2025-12-20 16:50:29,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.851840 seconds
+[2025-12-20 16:50:29,732][openai._base_client][INFO] - Retrying request to /chat/completions in 0.867341 seconds
+[2025-12-20 16:50:31,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:32,197][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:33,490][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:35,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:37,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:38,860][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:46,175][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:47,713][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:50:58,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:51:13,903][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:51:32,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:51:38,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494211 seconds
+[2025-12-20 16:51:43,562][openai._base_client][INFO] - Retrying request to /chat/completions in 0.782604 seconds
+[2025-12-20 16:52:24,525][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:24,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389727 seconds
+[2025-12-20 16:52:24,688][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443988 seconds
+[2025-12-20 16:52:28,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:29,587][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421516 seconds
+[2025-12-20 16:52:29,587][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430396 seconds
+[2025-12-20 16:52:29,588][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461349 seconds
+[2025-12-20 16:52:29,588][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405125 seconds
+[2025-12-20 16:52:29,976][openai._base_client][INFO] - Retrying request to /chat/completions in 0.817528 seconds
+[2025-12-20 16:52:31,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:32,368][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:32,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:35,014][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955816 seconds
+[2025-12-20 16:52:35,019][openai._base_client][INFO] - Retrying request to /chat/completions in 0.811696 seconds
+[2025-12-20 16:52:35,051][openai._base_client][INFO] - Retrying request to /chat/completions in 0.962378 seconds
+[2025-12-20 16:52:35,078][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:37,341][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:40,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:40,948][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:42,915][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:45,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:51,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:51,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:52:54,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:53:19,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:53:21,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:53:39,495][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:53:44,826][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407266 seconds
+[2025-12-20 16:54:24,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:24,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393049 seconds
+[2025-12-20 16:54:29,817][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482513 seconds
+[2025-12-20 16:54:29,818][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376206 seconds
+[2025-12-20 16:54:29,818][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441459 seconds
+[2025-12-20 16:54:29,818][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478526 seconds
+[2025-12-20 16:54:29,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404263 seconds
+[2025-12-20 16:54:29,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418580 seconds
+[2025-12-20 16:54:29,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462218 seconds
+[2025-12-20 16:54:29,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496087 seconds
+[2025-12-20 16:54:29,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449740 seconds
+[2025-12-20 16:54:29,819][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459502 seconds
+[2025-12-20 16:54:29,820][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459100 seconds
+[2025-12-20 16:54:29,820][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444177 seconds
+[2025-12-20 16:54:29,820][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377952 seconds
+[2025-12-20 16:54:29,820][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440545 seconds
+[2025-12-20 16:54:29,820][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467144 seconds
+[2025-12-20 16:54:30,208][openai._base_client][INFO] - Retrying request to /chat/completions in 0.848438 seconds
+[2025-12-20 16:54:34,470][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:35,197][openai._base_client][INFO] - Retrying request to /chat/completions in 0.936385 seconds
+[2025-12-20 16:54:35,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.926985 seconds
+[2025-12-20 16:54:35,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.890080 seconds
+[2025-12-20 16:54:37,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:38,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:39,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:39,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:39,906][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:42,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:44,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:44,794][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:45,299][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:46,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:46,895][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:48,874][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:49,472][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:54:59,181][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:04,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:05,151][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484550 seconds
+[2025-12-20 16:55:09,219][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:10,153][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419393 seconds
+[2025-12-20 16:55:10,154][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483059 seconds
+[2025-12-20 16:55:10,154][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449101 seconds
+[2025-12-20 16:55:10,638][openai._base_client][INFO] - Retrying request to /chat/completions in 0.929182 seconds
+[2025-12-20 16:55:11,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:12,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:12,512][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:13,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:14,304][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:15,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:15,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:15,606][openai._base_client][INFO] - Retrying request to /chat/completions in 0.793440 seconds
+[2025-12-20 16:55:15,639][openai._base_client][INFO] - Retrying request to /chat/completions in 0.796386 seconds
+[2025-12-20 16:55:17,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:20,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:36,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:40,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:41,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:43,204][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:44,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 16:55:49,414][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458191 seconds
+[2025-12-20 16:55:52,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-20/16-58-32/.hydra/hydra.yaml b/outputs/2025-12-20/16-58-32/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5dc0c39562c2d60aef9530fe6960967a7954821c
--- /dev/null
+++ b/outputs/2025-12-20/16-58-32/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/16-58-32
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/16-58-32/.hydra/overrides.yaml b/outputs/2025-12-20/16-58-32/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/16-58-32/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/16-58-32/eval_api.log b/outputs/2025-12-20/16-58-32/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..79f826816d2c7fc4b205a657078b56df6cdd367e
--- /dev/null
+++ b/outputs/2025-12-20/16-58-32/eval_api.log
@@ -0,0 +1,21 @@
+[2025-12-20 16:58:33,423][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,457][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,461][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,463][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,698][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,698][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,699][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,712][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,712][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,718][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:33,737][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:38,223][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448917 seconds
+[2025-12-20 16:58:38,224][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403869 seconds
+[2025-12-20 16:58:38,224][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413831 seconds
+[2025-12-20 16:58:38,225][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466682 seconds
+[2025-12-20 16:58:38,225][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452684 seconds
+[2025-12-20 16:58:38,643][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:38,653][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:38,687][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:38,692][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
+[2025-12-20 16:58:38,706][httpx][INFO] - HTTP Request: POST https://agentrouter.org/v1/chat/completions "HTTP/1.1 401 Unauthorized"
diff --git a/outputs/2025-12-20/17-01-22/.hydra/config.yaml b/outputs/2025-12-20/17-01-22/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f4a9bf0672043a8b9f07246b2e811f1493f167d9
--- /dev/null
+++ b/outputs/2025-12-20/17-01-22/.hydra/config.yaml
@@ -0,0 +1,923 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: null
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/qwen3-235b-a22b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  fireworks/models/qwen3-235b-a22b:
+    provider_name: openai
+    model_name: fireworks/models/qwen3-235b-a22b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/17-01-22/.hydra/overrides.yaml b/outputs/2025-12-20/17-01-22/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/17-01-22/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/17-01-22/eval_api.log b/outputs/2025-12-20/17-01-22/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-20/17-01-41/eval_api.log b/outputs/2025-12-20/17-01-41/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..73d9ce2dea25944a5c90c4c90ff7b49a8573ec6c
--- /dev/null
+++ b/outputs/2025-12-20/17-01-41/eval_api.log
@@ -0,0 +1,137 @@
+[2025-12-20 17:01:47,713][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488483 seconds
+[2025-12-20 17:01:47,714][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407279 seconds
+[2025-12-20 17:01:47,714][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443876 seconds
+[2025-12-20 17:01:47,715][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431324 seconds
+[2025-12-20 17:01:47,715][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460108 seconds
+[2025-12-20 17:01:47,715][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460191 seconds
+[2025-12-20 17:01:47,715][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392812 seconds
+[2025-12-20 17:01:47,715][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394182 seconds
+[2025-12-20 17:01:47,716][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414679 seconds
+[2025-12-20 17:01:47,716][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404355 seconds
+[2025-12-20 17:01:47,976][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424768 seconds
+[2025-12-20 17:01:49,275][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:01:49,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:01:49,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:01:49,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:01:50,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:01:53,113][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812977 seconds
+[2025-12-20 17:01:53,115][openai._base_client][INFO] - Retrying request to /chat/completions in 0.824005 seconds
+[2025-12-20 17:01:53,122][openai._base_client][INFO] - Retrying request to /chat/completions in 0.856364 seconds
+[2025-12-20 17:01:53,124][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864104 seconds
+[2025-12-20 17:01:53,132][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878456 seconds
+[2025-12-20 17:01:53,149][openai._base_client][INFO] - Retrying request to /chat/completions in 0.889136 seconds
+[2025-12-20 17:01:53,161][openai._base_client][INFO] - Retrying request to /chat/completions in 0.859499 seconds
+[2025-12-20 17:01:53,178][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904616 seconds
+[2025-12-20 17:01:53,178][openai._base_client][INFO] - Retrying request to /chat/completions in 0.836612 seconds
+[2025-12-20 17:01:53,206][openai._base_client][INFO] - Retrying request to /chat/completions in 0.752815 seconds
+[2025-12-20 17:01:53,403][openai._base_client][INFO] - Retrying request to /chat/completions in 0.827162 seconds
+[2025-12-20 17:02:00,068][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:00,078][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:00,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:00,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:00,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:03,592][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:08,602][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395286 seconds
+[2025-12-20 17:02:13,619][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440952 seconds
+[2025-12-20 17:02:13,620][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406958 seconds
+[2025-12-20 17:02:13,620][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387490 seconds
+[2025-12-20 17:02:13,620][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465612 seconds
+[2025-12-20 17:02:13,621][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416163 seconds
+[2025-12-20 17:02:13,878][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495653 seconds
+[2025-12-20 17:02:17,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:17,299][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:17,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:17,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:17,471][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:17,498][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:17,731][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:17,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:18,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:18,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:19,011][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871348 seconds
+[2025-12-20 17:02:22,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:22,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:22,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:22,124][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:22,241][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:27,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:32,288][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415653 seconds
+[2025-12-20 17:02:37,303][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492757 seconds
+[2025-12-20 17:02:37,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403039 seconds
+[2025-12-20 17:02:37,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487488 seconds
+[2025-12-20 17:02:37,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472986 seconds
+[2025-12-20 17:02:37,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435059 seconds
+[2025-12-20 17:02:37,709][openai._base_client][INFO] - Retrying request to /chat/completions in 0.884271 seconds
+[2025-12-20 17:02:39,724][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:39,931][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:40,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:40,108][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:40,116][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:40,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:40,307][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:40,360][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:40,598][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:42,712][openai._base_client][INFO] - Retrying request to /chat/completions in 0.882135 seconds
+[2025-12-20 17:02:42,781][openai._base_client][INFO] - Retrying request to /chat/completions in 0.847778 seconds
+[2025-12-20 17:02:42,794][openai._base_client][INFO] - Retrying request to /chat/completions in 0.810413 seconds
+[2025-12-20 17:02:43,831][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:45,141][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:45,643][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:49,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:49,435][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:49,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:02:54,449][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483085 seconds
+[2025-12-20 17:02:59,467][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420402 seconds
+[2025-12-20 17:02:59,468][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485535 seconds
+[2025-12-20 17:02:59,469][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382028 seconds
+[2025-12-20 17:02:59,469][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487399 seconds
+[2025-12-20 17:02:59,935][openai._base_client][INFO] - Retrying request to /chat/completions in 0.884555 seconds
+[2025-12-20 17:03:01,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:01,762][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:01,774][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:01,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:01,913][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:01,942][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:01,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:01,972][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:02,092][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:02,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:02,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:04,856][openai._base_client][INFO] - Retrying request to /chat/completions in 0.800210 seconds
+[2025-12-20 17:03:04,890][openai._base_client][INFO] - Retrying request to /chat/completions in 0.859649 seconds
+[2025-12-20 17:03:04,956][openai._base_client][INFO] - Retrying request to /chat/completions in 0.852976 seconds
+[2025-12-20 17:03:06,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:08,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:11,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:11,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:11,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:16,956][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396103 seconds
+[2025-12-20 17:03:21,975][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396281 seconds
+[2025-12-20 17:03:21,976][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468898 seconds
+[2025-12-20 17:03:21,977][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375767 seconds
+[2025-12-20 17:03:21,977][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465528 seconds
+[2025-12-20 17:03:21,977][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479948 seconds
+[2025-12-20 17:03:21,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459219 seconds
+[2025-12-20 17:03:21,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493713 seconds
+[2025-12-20 17:03:21,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475128 seconds
+[2025-12-20 17:03:21,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445899 seconds
+[2025-12-20 17:03:21,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387122 seconds
+[2025-12-20 17:03:21,979][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423877 seconds
+[2025-12-20 17:03:21,979][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375894 seconds
+[2025-12-20 17:03:22,356][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972964 seconds
+[2025-12-20 17:03:22,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:22,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:23,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:27,361][openai._base_client][INFO] - Retrying request to /chat/completions in 0.814692 seconds
+[2025-12-20 17:03:27,367][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933959 seconds
+[2025-12-20 17:03:27,404][openai._base_client][INFO] - Retrying request to /chat/completions in 0.828144 seconds
+[2025-12-20 17:03:27,426][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788748 seconds
+[2025-12-20 17:03:27,438][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774072 seconds
+[2025-12-20 17:03:27,444][openai._base_client][INFO] - Retrying request to /chat/completions in 0.835256 seconds
+[2025-12-20 17:03:27,455][openai._base_client][INFO] - Retrying request to /chat/completions in 0.984024 seconds
+[2025-12-20 17:03:27,460][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807014 seconds
+[2025-12-20 17:03:27,474][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878915 seconds
+[2025-12-20 17:03:28,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:28,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:28,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:03:29,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-20/17-04-07/eval_api.log b/outputs/2025-12-20/17-04-07/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..8abe8e177f6776d3b8581049555c4b76df6a5771
--- /dev/null
+++ b/outputs/2025-12-20/17-04-07/eval_api.log
@@ -0,0 +1,98 @@
+[2025-12-20 17:04:13,303][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429971 seconds
+[2025-12-20 17:04:13,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471568 seconds
+[2025-12-20 17:04:13,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469838 seconds
+[2025-12-20 17:04:13,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378903 seconds
+[2025-12-20 17:04:13,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415492 seconds
+[2025-12-20 17:04:13,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383240 seconds
+[2025-12-20 17:04:13,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459360 seconds
+[2025-12-20 17:04:13,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488121 seconds
+[2025-12-20 17:04:13,306][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420707 seconds
+[2025-12-20 17:04:13,306][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393051 seconds
+[2025-12-20 17:04:13,306][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495173 seconds
+[2025-12-20 17:04:13,306][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397040 seconds
+[2025-12-20 17:04:13,306][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494046 seconds
+[2025-12-20 17:04:13,306][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464509 seconds
+[2025-12-20 17:04:13,307][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423798 seconds
+[2025-12-20 17:04:13,307][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447601 seconds
+[2025-12-20 17:04:18,691][openai._base_client][INFO] - Retrying request to /chat/completions in 0.882112 seconds
+[2025-12-20 17:04:18,692][openai._base_client][INFO] - Retrying request to /chat/completions in 0.989096 seconds
+[2025-12-20 17:04:18,700][openai._base_client][INFO] - Retrying request to /chat/completions in 0.953222 seconds
+[2025-12-20 17:04:18,705][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788961 seconds
+[2025-12-20 17:04:18,722][openai._base_client][INFO] - Retrying request to /chat/completions in 0.940184 seconds
+[2025-12-20 17:04:18,728][openai._base_client][INFO] - Retrying request to /chat/completions in 0.971790 seconds
+[2025-12-20 17:04:18,732][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871945 seconds
+[2025-12-20 17:04:18,736][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755207 seconds
+[2025-12-20 17:04:18,756][openai._base_client][INFO] - Retrying request to /chat/completions in 0.877482 seconds
+[2025-12-20 17:04:18,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.756461 seconds
+[2025-12-20 17:04:18,773][openai._base_client][INFO] - Retrying request to /chat/completions in 0.834979 seconds
+[2025-12-20 17:04:18,776][openai._base_client][INFO] - Retrying request to /chat/completions in 0.840360 seconds
+[2025-12-20 17:04:18,778][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803762 seconds
+[2025-12-20 17:04:18,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.804042 seconds
+[2025-12-20 17:04:18,802][openai._base_client][INFO] - Retrying request to /chat/completions in 0.919801 seconds
+[2025-12-20 17:04:18,803][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955849 seconds
+[2025-12-20 17:04:34,785][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437880 seconds
+[2025-12-20 17:04:34,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446537 seconds
+[2025-12-20 17:04:34,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494980 seconds
+[2025-12-20 17:04:34,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498744 seconds
+[2025-12-20 17:04:34,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418681 seconds
+[2025-12-20 17:04:34,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392174 seconds
+[2025-12-20 17:04:34,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497924 seconds
+[2025-12-20 17:04:34,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471154 seconds
+[2025-12-20 17:04:34,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392635 seconds
+[2025-12-20 17:04:34,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493071 seconds
+[2025-12-20 17:04:34,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460487 seconds
+[2025-12-20 17:04:34,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419275 seconds
+[2025-12-20 17:04:34,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488907 seconds
+[2025-12-20 17:04:34,789][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406109 seconds
+[2025-12-20 17:04:34,789][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461068 seconds
+[2025-12-20 17:04:34,789][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387756 seconds
+[2025-12-20 17:04:40,183][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905208 seconds
+[2025-12-20 17:04:40,184][openai._base_client][INFO] - Retrying request to /chat/completions in 0.928989 seconds
+[2025-12-20 17:04:40,185][openai._base_client][INFO] - Retrying request to /chat/completions in 0.889791 seconds
+[2025-12-20 17:04:40,197][openai._base_client][INFO] - Retrying request to /chat/completions in 0.804760 seconds
+[2025-12-20 17:04:40,207][openai._base_client][INFO] - Retrying request to /chat/completions in 0.931564 seconds
+[2025-12-20 17:04:40,209][openai._base_client][INFO] - Retrying request to /chat/completions in 0.861961 seconds
+[2025-12-20 17:04:40,226][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803328 seconds
+[2025-12-20 17:04:40,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.919038 seconds
+[2025-12-20 17:04:40,251][openai._base_client][INFO] - Retrying request to /chat/completions in 0.768086 seconds
+[2025-12-20 17:04:40,251][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878961 seconds
+[2025-12-20 17:04:40,262][openai._base_client][INFO] - Retrying request to /chat/completions in 0.765498 seconds
+[2025-12-20 17:04:40,279][openai._base_client][INFO] - Retrying request to /chat/completions in 0.853278 seconds
+[2025-12-20 17:04:40,283][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860110 seconds
+[2025-12-20 17:04:40,283][openai._base_client][INFO] - Retrying request to /chat/completions in 0.789002 seconds
+[2025-12-20 17:04:40,287][openai._base_client][INFO] - Retrying request to /chat/completions in 0.953162 seconds
+[2025-12-20 17:04:40,287][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842310 seconds
+[2025-12-20 17:04:59,944][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:02,722][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:12,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:12,192][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:12,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:12,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:13,337][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:13,812][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:13,814][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:13,912][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:14,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:14,619][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:16,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:26,660][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:31,670][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377242 seconds
+[2025-12-20 17:05:36,684][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437739 seconds
+[2025-12-20 17:05:36,685][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399278 seconds
+[2025-12-20 17:05:36,685][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449046 seconds
+[2025-12-20 17:05:36,685][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493901 seconds
+[2025-12-20 17:05:36,686][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471682 seconds
+[2025-12-20 17:05:37,052][openai._base_client][INFO] - Retrying request to /chat/completions in 0.995670 seconds
+[2025-12-20 17:05:42,091][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870677 seconds
+[2025-12-20 17:05:42,125][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767332 seconds
+[2025-12-20 17:05:42,137][openai._base_client][INFO] - Retrying request to /chat/completions in 0.856232 seconds
+[2025-12-20 17:05:42,158][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944227 seconds
+[2025-12-20 17:05:42,181][openai._base_client][INFO] - Retrying request to /chat/completions in 0.942638 seconds
+[2025-12-20 17:05:54,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:54,700][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:54,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:55,593][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:56,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:56,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:05:57,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:06:00,733][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-20/17-06-28/.hydra/config.yaml b/outputs/2025-12-20/17-06-28/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1f38e389a4cdf58299e1b04558b32a1bbc34dde9
--- /dev/null
+++ b/outputs/2025-12-20/17-06-28/.hydra/config.yaml
@@ -0,0 +1,923 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 5
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/qwen3-235b-a22b
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  fireworks/models/qwen3-235b-a22b:
+    provider_name: openai
+    model_name: fireworks/models/qwen3-235b-a22b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  glm-4.6:
+    provider_name: openai
+    model_name: glm-4.6
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 7200
diff --git a/outputs/2025-12-20/17-08-03/.hydra/hydra.yaml b/outputs/2025-12-20/17-08-03/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f7613361d0cad46f087446a523f7c1d103d4dde
--- /dev/null
+++ b/outputs/2025-12-20/17-08-03/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-20/17-08-03
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-20/17-08-03/.hydra/overrides.yaml b/outputs/2025-12-20/17-08-03/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-20/17-08-03/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-20/17-17-00/eval_api.log b/outputs/2025-12-20/17-17-00/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..51852240e6494c3e570738b9a9709fbf7a0249f6
--- /dev/null
+++ b/outputs/2025-12-20/17-17-00/eval_api.log
@@ -0,0 +1,23 @@
+[2025-12-20 17:17:38,631][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:41,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:48,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:50,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:50,977][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:52,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:52,870][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:53,968][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:54,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:54,313][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:54,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:54,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:54,856][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:54,857][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:17:54,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:18:01,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-20 17:18:06,357][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418851 seconds
+[2025-12-20 17:18:11,371][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434416 seconds
+[2025-12-20 17:18:11,372][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390571 seconds
+[2025-12-20 17:18:11,778][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788888 seconds
+[2025-12-20 17:18:16,770][openai._base_client][INFO] - Retrying request to /chat/completions in 0.753281 seconds
+[2025-12-20 17:18:16,809][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972656 seconds
+[2025-12-20 17:18:19,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/00-59-27/.hydra/overrides.yaml b/outputs/2025-12-21/00-59-27/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/00-59-27/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/01-00-58/.hydra/overrides.yaml b/outputs/2025-12-21/01-00-58/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/01-00-58/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/01-00-58/eval_api.log b/outputs/2025-12-21/01-00-58/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..9499fe4967978d771fa95d650047849a833edcb3
--- /dev/null
+++ b/outputs/2025-12-21/01-00-58/eval_api.log
@@ -0,0 +1,31 @@
+[2025-12-21 01:01:01,833][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:01,852][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422100 seconds
+[2025-12-21 01:01:03,023][openai._base_client][INFO] - Retrying request to /chat/completions in 0.968952 seconds
+[2025-12-21 01:01:10,292][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:10,298][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463766 seconds
+[2025-12-21 01:01:15,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:15,164][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485795 seconds
+[2025-12-21 01:01:18,177][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:18,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395385 seconds
+[2025-12-21 01:01:27,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:27,019][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498569 seconds
+[2025-12-21 01:01:31,750][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:31,759][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480659 seconds
+[2025-12-21 01:01:35,864][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:35,873][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447854 seconds
+[2025-12-21 01:01:42,857][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:42,864][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395894 seconds
+[2025-12-21 01:01:53,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:53,274][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413905 seconds
+[2025-12-21 01:01:54,426][openai._base_client][INFO] - Retrying request to /chat/completions in 0.862500 seconds
+[2025-12-21 01:01:59,873][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:01:59,881][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386471 seconds
+[2025-12-21 01:02:07,378][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:02:07,387][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448051 seconds
+[2025-12-21 01:02:08,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.789655 seconds
+[2025-12-21 01:02:15,866][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455728 seconds
+[2025-12-21 01:02:17,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.982603 seconds
+[2025-12-21 01:02:28,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:02:33,710][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389627 seconds
+[2025-12-21 01:02:45,303][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:02:50,313][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437944 seconds
diff --git a/outputs/2025-12-21/01-07-17/eval_api.log b/outputs/2025-12-21/01-07-17/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..482326eaff8a2f4b9a12726268a6ef08eff4f8ca
--- /dev/null
+++ b/outputs/2025-12-21/01-07-17/eval_api.log
@@ -0,0 +1,4 @@
+[2025-12-21 01:07:23,118][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395831 seconds
+[2025-12-21 01:08:00,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:08:05,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383149 seconds
+[2025-12-21 01:08:43,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/01-10-11/.hydra/config.yaml b/outputs/2025-12-21/01-10-11/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..99dd7fac2a27f57f241d4ab1d62caed5a533b30c
--- /dev/null
+++ b/outputs/2025-12-21/01-10-11/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 1
+  max_turn: 100
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/kimi-k2-instruct
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  fireworks/models/kimi-k2-instruct:
+    provider_name: openai
+    model_name: fireworks/models/kimi-k2-instruct
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/01-10-11/.hydra/hydra.yaml b/outputs/2025-12-21/01-10-11/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b3065142cc4fc7d946f863ab2220945625a78f3d
--- /dev/null
+++ b/outputs/2025-12-21/01-10-11/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/01-10-11
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/01-10-11/.hydra/overrides.yaml b/outputs/2025-12-21/01-10-11/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/01-10-11/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/01-11-46/.hydra/config.yaml b/outputs/2025-12-21/01-11-46/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dfa27c7330e64bb5659dd8a58c7aa55b7215cea9
--- /dev/null
+++ b/outputs/2025-12-21/01-11-46/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 1
+  max_turn: 100
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/kimi-k2-instruct-0905
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  fireworks/models/kimi-k2-instruct-0905:
+    provider_name: openai
+    model_name: fireworks/models/kimi-k2-instruct-0905
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/01-11-46/.hydra/overrides.yaml b/outputs/2025-12-21/01-11-46/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/01-11-46/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/01-11-46/eval_api.log b/outputs/2025-12-21/01-11-46/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..1c43156de59ab57736874285e26a4659d72692f8
--- /dev/null
+++ b/outputs/2025-12-21/01-11-46/eval_api.log
@@ -0,0 +1,21 @@
+[2025-12-21 01:11:50,855][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:11:50,869][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477186 seconds
+[2025-12-21 01:11:53,573][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:11:53,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421610 seconds
+[2025-12-21 01:11:57,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:11:57,901][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395562 seconds
+[2025-12-21 01:12:03,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.760339 seconds
+[2025-12-21 01:12:16,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:12:16,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491398 seconds
+[2025-12-21 01:12:22,981][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:12:22,997][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426437 seconds
+[2025-12-21 01:12:26,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:12:26,449][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401722 seconds
+[2025-12-21 01:12:31,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:12:31,849][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495697 seconds
+[2025-12-21 01:12:39,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:12:39,243][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463501 seconds
+[2025-12-21 01:12:44,714][openai._base_client][INFO] - Retrying request to /chat/completions in 0.877483 seconds
+[2025-12-21 01:12:48,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 01:12:48,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411761 seconds
+[2025-12-21 01:12:54,370][openai._base_client][INFO] - Retrying request to /chat/completions in 0.855827 seconds
diff --git a/outputs/2025-12-21/01-13-13/.hydra/config.yaml b/outputs/2025-12-21/01-13-13/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fb00fd86b313314cb5d24a27b3dfc794257ff51b
--- /dev/null
+++ b/outputs/2025-12-21/01-13-13/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 1
+  max_turn: 400
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/kimi-k2-instruct-0905
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  fireworks/models/kimi-k2-instruct-0905:
+    provider_name: openai
+    model_name: fireworks/models/kimi-k2-instruct-0905
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/08-09-10/.hydra/config.yaml b/outputs/2025-12-21/08-09-10/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..dfa27c7330e64bb5659dd8a58c7aa55b7215cea9
--- /dev/null
+++ b/outputs/2025-12-21/08-09-10/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 1
+  max_turn: 100
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/kimi-k2-instruct-0905
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  fireworks/models/kimi-k2-instruct-0905:
+    provider_name: openai
+    model_name: fireworks/models/kimi-k2-instruct-0905
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/08-09-10/.hydra/hydra.yaml b/outputs/2025-12-21/08-09-10/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..225656bc68dcf310b459d1a1a6fc0dc38f04e657
--- /dev/null
+++ b/outputs/2025-12-21/08-09-10/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/08-09-10
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/08-42-41/.hydra/overrides.yaml b/outputs/2025-12-21/08-42-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..27b6904204a099b9569c74c6aaa2ca5ebf09f06d
--- /dev/null
+++ b/outputs/2025-12-21/08-42-41/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_15b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-1.5B-text_with_coordinates-qwen/global_step_1335
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-3-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-21/08-42-41/train.log b/outputs/2025-12-21/08-42-41/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-21/08-51-07/.hydra/config.yaml b/outputs/2025-12-21/08-51-07/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fc57677d92ff994bce122e2482de03ac92822204
--- /dev/null
+++ b/outputs/2025-12-21/08-51-07/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 100
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/kimi-k2-instruct-0905
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  fireworks/models/kimi-k2-instruct-0905:
+    provider_name: openai
+    model_name: fireworks/models/kimi-k2-instruct-0905
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/08-51-07/.hydra/hydra.yaml b/outputs/2025-12-21/08-51-07/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..287603eb286e0035171c4297e14335fdadb0afbc
--- /dev/null
+++ b/outputs/2025-12-21/08-51-07/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/08-51-07
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/08-51-07/.hydra/overrides.yaml b/outputs/2025-12-21/08-51-07/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/08-51-07/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/08-51-07/eval_api.log b/outputs/2025-12-21/08-51-07/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..c750cf28c021e488909697d09f9ebdeab1176d63
--- /dev/null
+++ b/outputs/2025-12-21/08-51-07/eval_api.log
@@ -0,0 +1,63 @@
+[2025-12-21 08:51:16,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:51:16,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453528 seconds
+[2025-12-21 08:51:21,970][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806594 seconds
+[2025-12-21 08:51:24,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:51:24,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414284 seconds
+[2025-12-21 08:51:26,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:51:26,494][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442054 seconds
+[2025-12-21 08:51:29,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:51:29,052][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442867 seconds
+[2025-12-21 08:51:34,502][openai._base_client][INFO] - Retrying request to /chat/completions in 0.945831 seconds
+[2025-12-21 08:51:36,630][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:51:36,640][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425097 seconds
+[2025-12-21 08:51:42,073][openai._base_client][INFO] - Retrying request to /chat/completions in 0.907242 seconds
+[2025-12-21 08:51:58,002][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390303 seconds
+[2025-12-21 08:52:00,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:00,110][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412533 seconds
+[2025-12-21 08:52:01,727][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:01,746][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493472 seconds
+[2025-12-21 08:52:07,246][openai._base_client][INFO] - Retrying request to /chat/completions in 0.800753 seconds
+[2025-12-21 08:52:15,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:15,389][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403348 seconds
+[2025-12-21 08:52:17,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:17,103][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470885 seconds
+[2025-12-21 08:52:22,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.841216 seconds
+[2025-12-21 08:52:24,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:24,495][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404907 seconds
+[2025-12-21 08:52:26,154][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:26,171][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405757 seconds
+[2025-12-21 08:52:28,091][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:28,108][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419344 seconds
+[2025-12-21 08:52:29,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:29,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423494 seconds
+[2025-12-21 08:52:31,468][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:31,491][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404867 seconds
+[2025-12-21 08:52:33,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:33,133][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467882 seconds
+[2025-12-21 08:52:34,805][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:34,827][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418795 seconds
+[2025-12-21 08:52:36,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:36,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386554 seconds
+[2025-12-21 08:52:38,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:38,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458198 seconds
+[2025-12-21 08:52:41,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:41,571][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388331 seconds
+[2025-12-21 08:52:44,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:44,118][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479701 seconds
+[2025-12-21 08:52:46,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:52:46,106][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431006 seconds
+[2025-12-21 08:52:51,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.992536 seconds
+[2025-12-21 08:53:04,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:53:04,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408494 seconds
+[2025-12-21 08:53:05,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:53:05,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483743 seconds
+[2025-12-21 08:53:08,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:53:08,618][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479368 seconds
+[2025-12-21 08:53:14,106][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755809 seconds
+[2025-12-21 08:53:16,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:53:16,350][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437255 seconds
+[2025-12-21 08:53:18,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:53:18,298][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486539 seconds
+[2025-12-21 08:53:23,791][openai._base_client][INFO] - Retrying request to /chat/completions in 0.789074 seconds
+[2025-12-21 08:53:36,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:53:36,147][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381616 seconds
diff --git a/outputs/2025-12-21/08-54-32/.hydra/config.yaml b/outputs/2025-12-21/08-54-32/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ee948561e675fb54b455870c91cdbdd1601a3b7e
--- /dev/null
+++ b/outputs/2025-12-21/08-54-32/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/kimi-k2-instruct-0905
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  fireworks/models/kimi-k2-instruct-0905:
+    provider_name: openai
+    model_name: fireworks/models/kimi-k2-instruct-0905
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/08-54-32/.hydra/hydra.yaml b/outputs/2025-12-21/08-54-32/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..3f404659c32c75bb60a34bd2098cbc49cc8b780e
--- /dev/null
+++ b/outputs/2025-12-21/08-54-32/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/08-54-32
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/08-54-32/eval_api.log b/outputs/2025-12-21/08-54-32/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..feb368264c5b5a5077c913bc23c49975dafa487a
--- /dev/null
+++ b/outputs/2025-12-21/08-54-32/eval_api.log
@@ -0,0 +1,1281 @@
+[2025-12-21 08:54:36,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:37,229][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:38,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:38,936][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384873 seconds
+[2025-12-21 08:54:38,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428756 seconds
+[2025-12-21 08:54:38,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448079 seconds
+[2025-12-21 08:54:38,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462391 seconds
+[2025-12-21 08:54:38,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471947 seconds
+[2025-12-21 08:54:38,938][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490558 seconds
+[2025-12-21 08:54:38,938][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391196 seconds
+[2025-12-21 08:54:38,938][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486348 seconds
+[2025-12-21 08:54:38,939][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471314 seconds
+[2025-12-21 08:54:38,939][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462270 seconds
+[2025-12-21 08:54:38,940][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424772 seconds
+[2025-12-21 08:54:39,813][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:42,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:42,207][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:43,011][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:43,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:44,325][openai._base_client][INFO] - Retrying request to /chat/completions in 0.802253 seconds
+[2025-12-21 08:54:44,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864390 seconds
+[2025-12-21 08:54:44,367][openai._base_client][INFO] - Retrying request to /chat/completions in 0.941411 seconds
+[2025-12-21 08:54:44,368][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788694 seconds
+[2025-12-21 08:54:44,388][openai._base_client][INFO] - Retrying request to /chat/completions in 0.762851 seconds
+[2025-12-21 08:54:44,403][openai._base_client][INFO] - Retrying request to /chat/completions in 0.866677 seconds
+[2025-12-21 08:54:44,403][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969169 seconds
+[2025-12-21 08:54:44,412][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967389 seconds
+[2025-12-21 08:54:44,412][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774130 seconds
+[2025-12-21 08:54:44,428][openai._base_client][INFO] - Retrying request to /chat/completions in 0.936034 seconds
+[2025-12-21 08:54:44,432][openai._base_client][INFO] - Retrying request to /chat/completions in 0.754911 seconds
+[2025-12-21 08:54:44,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:45,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:46,033][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:48,116][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:50,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:52,147][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:54,766][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:55,134][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408595 seconds
+[2025-12-21 08:54:55,157][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406400 seconds
+[2025-12-21 08:54:55,161][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389233 seconds
+[2025-12-21 08:54:55,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:57,757][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:57,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:57,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:54:58,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:07,654][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:12,664][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484369 seconds
+[2025-12-21 08:55:16,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:17,680][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496634 seconds
+[2025-12-21 08:55:17,681][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375637 seconds
+[2025-12-21 08:55:17,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449745 seconds
+[2025-12-21 08:55:17,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388549 seconds
+[2025-12-21 08:55:17,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462844 seconds
+[2025-12-21 08:55:17,683][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385837 seconds
+[2025-12-21 08:55:17,683][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383779 seconds
+[2025-12-21 08:55:17,684][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441898 seconds
+[2025-12-21 08:55:17,684][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497285 seconds
+[2025-12-21 08:55:18,153][openai._base_client][INFO] - Retrying request to /chat/completions in 0.886581 seconds
+[2025-12-21 08:55:23,073][openai._base_client][INFO] - Retrying request to /chat/completions in 0.859403 seconds
+[2025-12-21 08:55:23,074][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904037 seconds
+[2025-12-21 08:55:23,074][openai._base_client][INFO] - Retrying request to /chat/completions in 0.963632 seconds
+[2025-12-21 08:55:23,129][openai._base_client][INFO] - Retrying request to /chat/completions in 0.986463 seconds
+[2025-12-21 08:55:23,133][openai._base_client][INFO] - Retrying request to /chat/completions in 0.837734 seconds
+[2025-12-21 08:55:23,148][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949063 seconds
+[2025-12-21 08:55:23,180][openai._base_client][INFO] - Retrying request to /chat/completions in 0.951884 seconds
+[2025-12-21 08:55:23,183][openai._base_client][INFO] - Retrying request to /chat/completions in 0.952078 seconds
+[2025-12-21 08:55:23,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:28,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:34,146][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387413 seconds
+[2025-12-21 08:55:39,160][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375470 seconds
+[2025-12-21 08:55:39,161][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413177 seconds
+[2025-12-21 08:55:39,161][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489226 seconds
+[2025-12-21 08:55:39,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420149 seconds
+[2025-12-21 08:55:39,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471633 seconds
+[2025-12-21 08:55:39,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425912 seconds
+[2025-12-21 08:55:39,164][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392468 seconds
+[2025-12-21 08:55:39,538][openai._base_client][INFO] - Retrying request to /chat/completions in 0.928043 seconds
+[2025-12-21 08:55:44,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.787229 seconds
+[2025-12-21 08:55:44,559][openai._base_client][INFO] - Retrying request to /chat/completions in 0.839167 seconds
+[2025-12-21 08:55:44,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967494 seconds
+[2025-12-21 08:55:44,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.997775 seconds
+[2025-12-21 08:55:44,591][openai._base_client][INFO] - Retrying request to /chat/completions in 0.766256 seconds
+[2025-12-21 08:55:44,637][openai._base_client][INFO] - Retrying request to /chat/completions in 0.780521 seconds
+[2025-12-21 08:55:44,654][openai._base_client][INFO] - Retrying request to /chat/completions in 0.858834 seconds
+[2025-12-21 08:55:49,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:49,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:50,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:54,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:55:56,304][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:01,312][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442490 seconds
+[2025-12-21 08:56:03,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:05,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:06,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:06,090][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432213 seconds
+[2025-12-21 08:56:06,091][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431941 seconds
+[2025-12-21 08:56:06,093][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462964 seconds
+[2025-12-21 08:56:07,897][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:08,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:08,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:08,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:08,873][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:09,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:09,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:09,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:09,432][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:09,522][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:09,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:09,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:10,128][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:10,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:10,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:11,133][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:11,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:11,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:11,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:11,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:11,631][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:11,677][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:12,119][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:12,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:12,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:12,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:12,979][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:13,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:15,536][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:15,910][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:16,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:17,341][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:17,448][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463235 seconds
+[2025-12-21 08:56:17,449][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436365 seconds
+[2025-12-21 08:56:17,451][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384018 seconds
+[2025-12-21 08:56:17,452][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408529 seconds
+[2025-12-21 08:56:17,453][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479529 seconds
+[2025-12-21 08:56:17,454][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400299 seconds
+[2025-12-21 08:56:17,456][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386100 seconds
+[2025-12-21 08:56:17,458][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422118 seconds
+[2025-12-21 08:56:18,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:18,886][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:19,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:19,404][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:19,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:19,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:19,585][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:19,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:20,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:20,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:20,505][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:20,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:20,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:20,685][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:20,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:21,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,158][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,204][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,214][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,386][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:22,841][openai._base_client][INFO] - Retrying request to /chat/completions in 0.973585 seconds
+[2025-12-21 08:56:22,845][openai._base_client][INFO] - Retrying request to /chat/completions in 0.811261 seconds
+[2025-12-21 08:56:22,856][openai._base_client][INFO] - Retrying request to /chat/completions in 0.804590 seconds
+[2025-12-21 08:56:22,863][openai._base_client][INFO] - Retrying request to /chat/completions in 0.892186 seconds
+[2025-12-21 08:56:22,883][openai._base_client][INFO] - Retrying request to /chat/completions in 0.839176 seconds
+[2025-12-21 08:56:22,888][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901611 seconds
+[2025-12-21 08:56:22,914][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767757 seconds
+[2025-12-21 08:56:22,935][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863274 seconds
+[2025-12-21 08:56:24,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:24,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:24,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:25,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:25,270][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:25,310][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:25,758][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:25,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:25,875][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389775 seconds
+[2025-12-21 08:56:25,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403758 seconds
+[2025-12-21 08:56:25,879][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449056 seconds
+[2025-12-21 08:56:25,881][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422090 seconds
+[2025-12-21 08:56:25,882][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474195 seconds
+[2025-12-21 08:56:25,883][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393532 seconds
+[2025-12-21 08:56:25,885][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375692 seconds
+[2025-12-21 08:56:25,888][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457217 seconds
+[2025-12-21 08:56:30,890][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492297 seconds
+[2025-12-21 08:56:30,892][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458142 seconds
+[2025-12-21 08:56:30,892][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404301 seconds
+[2025-12-21 08:56:30,892][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385047 seconds
+[2025-12-21 08:56:30,893][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470882 seconds
+[2025-12-21 08:56:30,893][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458003 seconds
+[2025-12-21 08:56:30,893][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465015 seconds
+[2025-12-21 08:56:30,894][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451520 seconds
+[2025-12-21 08:56:31,265][openai._base_client][INFO] - Retrying request to /chat/completions in 0.798328 seconds
+[2025-12-21 08:56:31,266][openai._base_client][INFO] - Retrying request to /chat/completions in 0.924704 seconds
+[2025-12-21 08:56:31,279][openai._base_client][INFO] - Retrying request to /chat/completions in 0.772711 seconds
+[2025-12-21 08:56:31,282][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806651 seconds
+[2025-12-21 08:56:31,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979818 seconds
+[2025-12-21 08:56:31,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.976356 seconds
+[2025-12-21 08:56:31,348][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755291 seconds
+[2025-12-21 08:56:31,359][openai._base_client][INFO] - Retrying request to /chat/completions in 0.893724 seconds
+[2025-12-21 08:56:36,283][openai._base_client][INFO] - Retrying request to /chat/completions in 0.756636 seconds
+[2025-12-21 08:56:36,299][openai._base_client][INFO] - Retrying request to /chat/completions in 0.943310 seconds
+[2025-12-21 08:56:36,348][openai._base_client][INFO] - Retrying request to /chat/completions in 0.785681 seconds
+[2025-12-21 08:56:36,352][openai._base_client][INFO] - Retrying request to /chat/completions in 0.995823 seconds
+[2025-12-21 08:56:36,354][openai._base_client][INFO] - Retrying request to /chat/completions in 0.976542 seconds
+[2025-12-21 08:56:36,360][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774278 seconds
+[2025-12-21 08:56:36,366][openai._base_client][INFO] - Retrying request to /chat/completions in 0.778284 seconds
+[2025-12-21 08:56:36,387][openai._base_client][INFO] - Retrying request to /chat/completions in 0.866931 seconds
+[2025-12-21 08:56:42,059][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391432 seconds
+[2025-12-21 08:56:42,070][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419775 seconds
+[2025-12-21 08:56:42,094][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436887 seconds
+[2025-12-21 08:56:42,110][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400275 seconds
+[2025-12-21 08:56:42,197][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410723 seconds
+[2025-12-21 08:56:42,260][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398070 seconds
+[2025-12-21 08:56:42,291][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472061 seconds
+[2025-12-21 08:56:42,316][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435920 seconds
+[2025-12-21 08:56:47,055][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409507 seconds
+[2025-12-21 08:56:47,142][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490300 seconds
+[2025-12-21 08:56:47,142][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378061 seconds
+[2025-12-21 08:56:47,149][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493987 seconds
+[2025-12-21 08:56:47,248][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434635 seconds
+[2025-12-21 08:56:47,260][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396287 seconds
+[2025-12-21 08:56:47,337][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497706 seconds
+[2025-12-21 08:56:47,353][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387679 seconds
+[2025-12-21 08:56:47,454][openai._base_client][INFO] - Retrying request to /chat/completions in 0.841094 seconds
+[2025-12-21 08:56:47,493][openai._base_client][INFO] - Retrying request to /chat/completions in 0.907761 seconds
+[2025-12-21 08:56:47,514][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939453 seconds
+[2025-12-21 08:56:47,534][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818314 seconds
+[2025-12-21 08:56:47,610][openai._base_client][INFO] - Retrying request to /chat/completions in 0.845751 seconds
+[2025-12-21 08:56:47,663][openai._base_client][INFO] - Retrying request to /chat/completions in 0.772609 seconds
+[2025-12-21 08:56:47,754][openai._base_client][INFO] - Retrying request to /chat/completions in 0.787697 seconds
+[2025-12-21 08:56:47,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.850547 seconds
+[2025-12-21 08:56:49,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:49,719][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:49,788][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:50,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:51,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:51,135][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:51,181][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:51,703][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:51,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:52,248][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:52,468][openai._base_client][INFO] - Retrying request to /chat/completions in 0.780660 seconds
+[2025-12-21 08:56:52,524][openai._base_client][INFO] - Retrying request to /chat/completions in 0.913740 seconds
+[2025-12-21 08:56:53,112][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:54,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:54,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:54,998][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:56:59,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:04,835][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459329 seconds
+[2025-12-21 08:57:06,132][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:06,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:06,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:06,722][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:07,072][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:07,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:07,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:07,693][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:07,695][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:07,717][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:07,886][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:08,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:08,751][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:10,157][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:10,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:10,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:14,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:14,618][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390116 seconds
+[2025-12-21 08:57:14,619][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472631 seconds
+[2025-12-21 08:57:14,620][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424976 seconds
+[2025-12-21 08:57:14,623][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487963 seconds
+[2025-12-21 08:57:15,382][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803227 seconds
+[2025-12-21 08:57:15,418][openai._base_client][INFO] - Retrying request to /chat/completions in 0.931268 seconds
+[2025-12-21 08:57:15,467][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985900 seconds
+[2025-12-21 08:57:15,483][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774520 seconds
+[2025-12-21 08:57:15,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:16,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:16,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:17,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:17,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:18,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:18,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:18,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:18,448][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:18,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:18,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:18,842][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:19,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:19,383][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:19,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:19,691][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:19,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:19,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:19,885][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475099 seconds
+[2025-12-21 08:57:19,985][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:20,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:20,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:20,700][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:21,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:21,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:21,634][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401674 seconds
+[2025-12-21 08:57:21,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:21,727][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:22,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:23,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:28,048][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418011 seconds
+[2025-12-21 08:57:28,421][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452991 seconds
+[2025-12-21 08:57:29,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:30,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:33,058][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470687 seconds
+[2025-12-21 08:57:33,879][openai._base_client][INFO] - Retrying request to /chat/completions in 0.922010 seconds
+[2025-12-21 08:57:36,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:41,033][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:41,203][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479831 seconds
+[2025-12-21 08:57:41,205][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447143 seconds
+[2025-12-21 08:57:42,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:42,439][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:42,557][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:42,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:43,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:43,435][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:43,515][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:43,523][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:43,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,168][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,744][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,897][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:44,930][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:45,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:45,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:45,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:45,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:45,727][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:45,735][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:46,085][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:46,208][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455905 seconds
+[2025-12-21 08:57:46,209][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484267 seconds
+[2025-12-21 08:57:46,210][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385807 seconds
+[2025-12-21 08:57:46,210][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437675 seconds
+[2025-12-21 08:57:46,656][openai._base_client][INFO] - Retrying request to /chat/completions in 0.799111 seconds
+[2025-12-21 08:57:47,399][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:47,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:47,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:48,005][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:48,489][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:49,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:49,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:52,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:52,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450226 seconds
+[2025-12-21 08:57:52,282][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418821 seconds
+[2025-12-21 08:57:52,284][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493422 seconds
+[2025-12-21 08:57:52,285][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446044 seconds
+[2025-12-21 08:57:52,286][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484494 seconds
+[2025-12-21 08:57:52,287][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483039 seconds
+[2025-12-21 08:57:52,288][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382394 seconds
+[2025-12-21 08:57:52,723][openai._base_client][INFO] - Retrying request to /chat/completions in 0.927179 seconds
+[2025-12-21 08:57:52,785][openai._base_client][INFO] - Retrying request to /chat/completions in 0.772878 seconds
+[2025-12-21 08:57:53,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:53,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:53,657][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421910 seconds
+[2025-12-21 08:57:53,803][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:54,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:54,921][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:57,292][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497808 seconds
+[2025-12-21 08:57:57,302][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470998 seconds
+[2025-12-21 08:57:57,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383735 seconds
+[2025-12-21 08:57:57,703][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860117 seconds
+[2025-12-21 08:57:57,734][openai._base_client][INFO] - Retrying request to /chat/completions in 0.989477 seconds
+[2025-12-21 08:57:57,746][openai._base_client][INFO] - Retrying request to /chat/completions in 0.956336 seconds
+[2025-12-21 08:57:57,768][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:57,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:57,777][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955289 seconds
+[2025-12-21 08:57:57,777][openai._base_client][INFO] - Retrying request to /chat/completions in 0.768895 seconds
+[2025-12-21 08:57:57,780][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864090 seconds
+[2025-12-21 08:57:58,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:58,623][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:58,705][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380176 seconds
+[2025-12-21 08:57:59,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.903114 seconds
+[2025-12-21 08:57:59,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:59,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:57:59,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:00,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:01,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:01,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:02,780][openai._base_client][INFO] - Retrying request to /chat/completions in 0.862547 seconds
+[2025-12-21 08:58:02,803][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818929 seconds
+[2025-12-21 08:58:03,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:03,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:03,624][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:03,755][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477746 seconds
+[2025-12-21 08:58:04,087][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864720 seconds
+[2025-12-21 08:58:04,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:04,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:04,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:04,983][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:05,650][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:10,660][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458824 seconds
+[2025-12-21 08:58:10,744][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455188 seconds
+[2025-12-21 08:58:11,176][openai._base_client][INFO] - Retrying request to /chat/completions in 0.880066 seconds
+[2025-12-21 08:58:11,254][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899088 seconds
+[2025-12-21 08:58:15,696][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409067 seconds
+[2025-12-21 08:58:15,697][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399320 seconds
+[2025-12-21 08:58:15,697][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413979 seconds
+[2025-12-21 08:58:15,697][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485632 seconds
+[2025-12-21 08:58:15,697][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421376 seconds
+[2025-12-21 08:58:15,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404717 seconds
+[2025-12-21 08:58:15,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487671 seconds
+[2025-12-21 08:58:19,322][openai._base_client][INFO] - Retrying request to /chat/completions in 0.876680 seconds
+[2025-12-21 08:58:21,100][openai._base_client][INFO] - Retrying request to /chat/completions in 0.843813 seconds
+[2025-12-21 08:58:21,104][openai._base_client][INFO] - Retrying request to /chat/completions in 0.802353 seconds
+[2025-12-21 08:58:21,108][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812726 seconds
+[2025-12-21 08:58:21,112][openai._base_client][INFO] - Retrying request to /chat/completions in 0.877330 seconds
+[2025-12-21 08:58:21,121][openai._base_client][INFO] - Retrying request to /chat/completions in 0.902951 seconds
+[2025-12-21 08:58:21,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972041 seconds
+[2025-12-21 08:58:23,232][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:23,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:25,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:25,759][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:32,171][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485745 seconds
+[2025-12-21 08:58:33,500][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:34,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:34,282][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:35,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:37,185][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402816 seconds
+[2025-12-21 08:58:39,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:39,571][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491272 seconds
+[2025-12-21 08:58:39,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484114 seconds
+[2025-12-21 08:58:40,788][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:40,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:41,039][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:41,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:41,252][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:41,606][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:41,735][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:41,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:41,916][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:42,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:42,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:42,585][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:42,657][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:42,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:42,851][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:43,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:44,256][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:44,576][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415337 seconds
+[2025-12-21 08:58:44,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477020 seconds
+[2025-12-21 08:58:44,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421733 seconds
+[2025-12-21 08:58:44,579][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419881 seconds
+[2025-12-21 08:58:44,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:45,063][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871968 seconds
+[2025-12-21 08:58:45,068][openai._base_client][INFO] - Retrying request to /chat/completions in 0.760682 seconds
+[2025-12-21 08:58:45,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:45,967][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:46,476][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:47,166][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:47,219][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:48,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:48,571][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435278 seconds
+[2025-12-21 08:58:48,572][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386456 seconds
+[2025-12-21 08:58:48,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469653 seconds
+[2025-12-21 08:58:48,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377647 seconds
+[2025-12-21 08:58:48,579][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426490 seconds
+[2025-12-21 08:58:48,580][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468725 seconds
+[2025-12-21 08:58:48,581][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451076 seconds
+[2025-12-21 08:58:48,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417917 seconds
+[2025-12-21 08:58:48,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468624 seconds
+[2025-12-21 08:58:49,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:49,803][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:50,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:50,206][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:50,426][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:50,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:50,877][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:50,890][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:50,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:51,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:51,649][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:51,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:52,251][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:52,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:52,282][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:52,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:52,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:52,884][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:52,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:53,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:53,243][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:53,432][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:53,585][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473270 seconds
+[2025-12-21 08:58:53,586][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409404 seconds
+[2025-12-21 08:58:53,737][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:53,962][openai._base_client][INFO] - Retrying request to /chat/completions in 0.982642 seconds
+[2025-12-21 08:58:54,003][openai._base_client][INFO] - Retrying request to /chat/completions in 0.918924 seconds
+[2025-12-21 08:58:54,008][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999306 seconds
+[2025-12-21 08:58:54,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.826051 seconds
+[2025-12-21 08:58:54,048][openai._base_client][INFO] - Retrying request to /chat/completions in 0.839498 seconds
+[2025-12-21 08:58:54,052][openai._base_client][INFO] - Retrying request to /chat/completions in 0.881877 seconds
+[2025-12-21 08:58:54,056][openai._base_client][INFO] - Retrying request to /chat/completions in 0.853849 seconds
+[2025-12-21 08:58:55,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:56,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:56,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:58,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:58,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:58,375][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:58:59,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:05,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455072 seconds
+[2025-12-21 08:59:10,031][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414445 seconds
+[2025-12-21 08:59:10,481][openai._base_client][INFO] - Retrying request to /chat/completions in 0.930725 seconds
+[2025-12-21 08:59:15,454][openai._base_client][INFO] - Retrying request to /chat/completions in 0.954102 seconds
+[2025-12-21 08:59:31,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380966 seconds
+[2025-12-21 08:59:31,434][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452646 seconds
+[2025-12-21 08:59:33,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:33,351][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:33,555][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495092 seconds
+[2025-12-21 08:59:33,557][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408986 seconds
+[2025-12-21 08:59:34,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:34,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:34,870][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:34,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,222][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,241][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,665][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,680][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:35,952][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,254][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,262][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,557][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:36,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:37,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:37,131][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:39,264][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:39,522][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410337 seconds
+[2025-12-21 08:59:39,525][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408971 seconds
+[2025-12-21 08:59:39,528][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409960 seconds
+[2025-12-21 08:59:39,530][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431810 seconds
+[2025-12-21 08:59:39,531][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381956 seconds
+[2025-12-21 08:59:39,534][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408308 seconds
+[2025-12-21 08:59:39,536][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424264 seconds
+[2025-12-21 08:59:39,538][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415487 seconds
+[2025-12-21 08:59:39,539][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408505 seconds
+[2025-12-21 08:59:39,540][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476782 seconds
+[2025-12-21 08:59:39,541][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484770 seconds
+[2025-12-21 08:59:39,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380168 seconds
+[2025-12-21 08:59:39,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433302 seconds
+[2025-12-21 08:59:39,543][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498364 seconds
+[2025-12-21 08:59:39,544][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444359 seconds
+[2025-12-21 08:59:39,545][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408878 seconds
+[2025-12-21 08:59:41,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,254][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,450][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,571][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,587][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,606][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:41,920][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:42,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:42,914][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:42,933][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:42,953][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:42,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:42,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:42,998][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:43,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:44,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:44,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:45,428][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:45,444][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:45,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:45,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:45,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,743][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,779][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,786][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:46,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:47,092][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404204 seconds
+[2025-12-21 08:59:47,094][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405913 seconds
+[2025-12-21 08:59:47,097][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457960 seconds
+[2025-12-21 08:59:47,099][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453157 seconds
+[2025-12-21 08:59:47,101][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456749 seconds
+[2025-12-21 08:59:47,102][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477275 seconds
+[2025-12-21 08:59:47,104][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477024 seconds
+[2025-12-21 08:59:47,105][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406899 seconds
+[2025-12-21 08:59:47,106][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480263 seconds
+[2025-12-21 08:59:47,107][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436178 seconds
+[2025-12-21 08:59:47,108][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481151 seconds
+[2025-12-21 08:59:47,109][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476749 seconds
+[2025-12-21 08:59:47,110][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452882 seconds
+[2025-12-21 08:59:47,111][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404925 seconds
+[2025-12-21 08:59:47,111][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497371 seconds
+[2025-12-21 08:59:47,112][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419188 seconds
+[2025-12-21 08:59:48,779][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:48,847][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:48,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,107][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,119][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,144][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,161][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,901][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,909][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:49,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,154][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,172][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,182][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,584][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:50,719][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,477][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,509][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,530][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:51,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419777 seconds
+[2025-12-21 08:59:51,832][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391447 seconds
+[2025-12-21 08:59:51,837][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455140 seconds
+[2025-12-21 08:59:51,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451712 seconds
+[2025-12-21 08:59:51,843][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418066 seconds
+[2025-12-21 08:59:51,846][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459430 seconds
+[2025-12-21 08:59:51,848][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445550 seconds
+[2025-12-21 08:59:51,851][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394470 seconds
+[2025-12-21 08:59:51,853][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391046 seconds
+[2025-12-21 08:59:51,854][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450338 seconds
+[2025-12-21 08:59:51,856][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437618 seconds
+[2025-12-21 08:59:51,858][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496704 seconds
+[2025-12-21 08:59:51,859][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441125 seconds
+[2025-12-21 08:59:51,860][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392921 seconds
+[2025-12-21 08:59:51,862][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499664 seconds
+[2025-12-21 08:59:51,862][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480800 seconds
+[2025-12-21 08:59:53,479][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:53,498][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,737][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,749][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:54,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:55,384][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:55,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,128][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,146][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,163][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,182][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,666][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,733][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,751][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:56,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:57,065][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405120 seconds
+[2025-12-21 08:59:57,067][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449994 seconds
+[2025-12-21 08:59:57,070][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383478 seconds
+[2025-12-21 08:59:57,072][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486793 seconds
+[2025-12-21 08:59:57,074][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444396 seconds
+[2025-12-21 08:59:57,075][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485009 seconds
+[2025-12-21 08:59:57,077][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483463 seconds
+[2025-12-21 08:59:57,078][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381392 seconds
+[2025-12-21 08:59:57,079][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459583 seconds
+[2025-12-21 08:59:57,080][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382504 seconds
+[2025-12-21 08:59:57,081][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401605 seconds
+[2025-12-21 08:59:57,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382913 seconds
+[2025-12-21 08:59:57,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403571 seconds
+[2025-12-21 08:59:57,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488490 seconds
+[2025-12-21 08:59:57,084][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453465 seconds
+[2025-12-21 08:59:57,084][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435609 seconds
+[2025-12-21 08:59:58,843][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:58,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,066][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,088][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,271][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,827][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 08:59:59,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,756][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,774][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,786][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:00,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:01,511][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:02,068][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:02,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:02,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:02,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:02,131][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:02,521][openai._base_client][INFO] - Retrying request to /chat/completions in 0.886553 seconds
+[2025-12-21 09:00:02,562][openai._base_client][INFO] - Retrying request to /chat/completions in 0.825701 seconds
+[2025-12-21 09:00:02,564][openai._base_client][INFO] - Retrying request to /chat/completions in 0.813030 seconds
+[2025-12-21 09:00:02,575][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783099 seconds
+[2025-12-21 09:00:03,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:03,392][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:04,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:04,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:05,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:05,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:06,688][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:07,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479959 seconds
+[2025-12-21 09:00:07,015][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388398 seconds
+[2025-12-21 09:00:07,016][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484905 seconds
+[2025-12-21 09:00:07,018][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467668 seconds
+[2025-12-21 09:00:07,020][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402152 seconds
+[2025-12-21 09:00:07,021][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488898 seconds
+[2025-12-21 09:00:07,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385286 seconds
+[2025-12-21 09:00:07,024][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481938 seconds
+[2025-12-21 09:00:07,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483620 seconds
+[2025-12-21 09:00:08,372][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:08,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:08,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:08,892][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:08,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:08,919][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:09,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:09,650][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:09,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:10,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:10,337][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:10,358][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:10,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,091][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,696][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,704][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,754][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,810][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:11,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:12,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431276 seconds
+[2025-12-21 09:00:12,028][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382105 seconds
+[2025-12-21 09:00:12,029][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495551 seconds
+[2025-12-21 09:00:12,282][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:12,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:12,446][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:12,489][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901303 seconds
+[2025-12-21 09:00:12,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.948216 seconds
+[2025-12-21 09:00:12,511][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908203 seconds
+[2025-12-21 09:00:12,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:13,500][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:13,668][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:14,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:14,967][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:14,976][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:17,136][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:17,377][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497885 seconds
+[2025-12-21 09:00:17,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429809 seconds
+[2025-12-21 09:00:17,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481678 seconds
+[2025-12-21 09:00:17,467][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400308 seconds
+[2025-12-21 09:00:17,468][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405587 seconds
+[2025-12-21 09:00:17,469][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398392 seconds
+[2025-12-21 09:00:17,471][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401673 seconds
+[2025-12-21 09:00:19,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:20,257][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:20,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:20,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:20,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:20,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:20,470][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:20,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:21,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:21,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:21,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:21,847][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:21,859][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:22,163][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:22,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:22,471][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494424 seconds
+[2025-12-21 09:00:22,471][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487405 seconds
+[2025-12-21 09:00:22,472][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427928 seconds
+[2025-12-21 09:00:22,472][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392417 seconds
+[2025-12-21 09:00:22,473][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435012 seconds
+[2025-12-21 09:00:22,473][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410221 seconds
+[2025-12-21 09:00:22,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:22,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767562 seconds
+[2025-12-21 09:00:22,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.991740 seconds
+[2025-12-21 09:00:22,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.787581 seconds
+[2025-12-21 09:00:23,029][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:23,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:23,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:23,201][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:23,204][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:23,800][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:25,033][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:25,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:25,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:25,072][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:25,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:27,244][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:27,869][openai._base_client][INFO] - Retrying request to /chat/completions in 0.973673 seconds
+[2025-12-21 09:00:27,886][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969586 seconds
+[2025-12-21 09:00:30,132][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:30,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:30,717][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:30,734][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:31,107][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414456 seconds
+[2025-12-21 09:00:31,109][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431383 seconds
+[2025-12-21 09:00:31,110][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424585 seconds
+[2025-12-21 09:00:31,111][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483354 seconds
+[2025-12-21 09:00:31,113][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481039 seconds
+[2025-12-21 09:00:32,663][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:32,961][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:33,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:34,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:34,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,855][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,905][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,917][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:35,974][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:36,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:36,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:36,599][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905778 seconds
+[2025-12-21 09:00:36,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:36,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,310][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,329][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,357][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:37,407][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:38,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:40,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:40,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:40,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:40,512][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:40,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:40,837][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491633 seconds
+[2025-12-21 09:00:40,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419667 seconds
+[2025-12-21 09:00:40,841][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431176 seconds
+[2025-12-21 09:00:40,843][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446613 seconds
+[2025-12-21 09:00:40,845][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493664 seconds
+[2025-12-21 09:00:40,846][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452609 seconds
+[2025-12-21 09:00:40,848][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442086 seconds
+[2025-12-21 09:00:40,849][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401799 seconds
+[2025-12-21 09:00:40,850][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432035 seconds
+[2025-12-21 09:00:40,851][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448201 seconds
+[2025-12-21 09:00:40,852][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387256 seconds
+[2025-12-21 09:00:40,852][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471284 seconds
+[2025-12-21 09:00:40,853][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462695 seconds
+[2025-12-21 09:00:40,854][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454265 seconds
+[2025-12-21 09:00:40,854][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473187 seconds
+[2025-12-21 09:00:43,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,186][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,194][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,488][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,542][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:43,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,566][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,659][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:44,725][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,119][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:45,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:46,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:46,253][openai._base_client][INFO] - Retrying request to /chat/completions in 0.977957 seconds
+[2025-12-21 09:00:46,275][openai._base_client][INFO] - Retrying request to /chat/completions in 0.995791 seconds
+[2025-12-21 09:00:46,303][openai._base_client][INFO] - Retrying request to /chat/completions in 0.952602 seconds
+[2025-12-21 09:00:46,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.941615 seconds
+[2025-12-21 09:00:46,388][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:46,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:48,359][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:48,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:48,910][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:48,917][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:49,318][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462212 seconds
+[2025-12-21 09:00:49,320][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404955 seconds
+[2025-12-21 09:00:49,322][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416273 seconds
+[2025-12-21 09:00:49,324][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426389 seconds
+[2025-12-21 09:00:49,325][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400401 seconds
+[2025-12-21 09:00:49,327][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443889 seconds
+[2025-12-21 09:00:49,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421571 seconds
+[2025-12-21 09:00:49,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473301 seconds
+[2025-12-21 09:00:49,331][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451289 seconds
+[2025-12-21 09:00:49,333][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393725 seconds
+[2025-12-21 09:00:49,334][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489787 seconds
+[2025-12-21 09:00:49,335][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427443 seconds
+[2025-12-21 09:00:50,828][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:51,195][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:51,744][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:51,761][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,681][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,704][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,748][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:52,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:53,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:53,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,052][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,060][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,336][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383428 seconds
+[2025-12-21 09:00:54,337][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398819 seconds
+[2025-12-21 09:00:54,566][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,668][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:54,783][openai._base_client][INFO] - Retrying request to /chat/completions in 0.954664 seconds
+[2025-12-21 09:00:54,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.907348 seconds
+[2025-12-21 09:00:54,827][openai._base_client][INFO] - Retrying request to /chat/completions in 0.993883 seconds
+[2025-12-21 09:00:55,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:55,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:56,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:56,645][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:56,815][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:57,236][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:57,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:57,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:00:57,795][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427933 seconds
+[2025-12-21 09:00:57,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481203 seconds
+[2025-12-21 09:00:57,800][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421037 seconds
+[2025-12-21 09:00:57,802][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437239 seconds
+[2025-12-21 09:00:57,804][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459069 seconds
+[2025-12-21 09:00:57,806][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493108 seconds
+[2025-12-21 09:00:57,807][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498474 seconds
+[2025-12-21 09:00:57,808][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433896 seconds
+[2025-12-21 09:00:57,810][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458909 seconds
+[2025-12-21 09:00:57,811][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423037 seconds
+[2025-12-21 09:00:57,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382137 seconds
+[2025-12-21 09:01:00,482][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,615][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,664][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,717][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:00,810][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:01,794][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:01,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:01,979][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:01,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:02,815][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459899 seconds
+[2025-12-21 09:01:02,815][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469973 seconds
+[2025-12-21 09:01:02,818][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:02,884][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:02,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:02,913][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:02,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:03,200][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972160 seconds
+[2025-12-21 09:01:03,225][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899018 seconds
+[2025-12-21 09:01:03,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.853025 seconds
+[2025-12-21 09:01:03,242][openai._base_client][INFO] - Retrying request to /chat/completions in 0.882237 seconds
+[2025-12-21 09:01:03,267][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949281 seconds
+[2025-12-21 09:01:03,335][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:03,338][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:03,645][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:03,652][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:04,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:04,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:04,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:04,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:05,099][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:05,567][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:05,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:05,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:06,197][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:06,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:06,604][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449261 seconds
+[2025-12-21 09:01:06,605][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466041 seconds
+[2025-12-21 09:01:06,607][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414880 seconds
+[2025-12-21 09:01:06,608][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485019 seconds
+[2025-12-21 09:01:06,609][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396916 seconds
+[2025-12-21 09:01:06,610][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381168 seconds
+[2025-12-21 09:01:06,611][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375804 seconds
+[2025-12-21 09:01:06,612][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432829 seconds
+[2025-12-21 09:01:06,614][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412098 seconds
+[2025-12-21 09:01:08,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:08,549][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:08,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:08,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:08,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,090][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,131][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,151][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,691][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:09,700][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,114][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,265][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,747][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,760][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,775][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,879][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:10,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:11,371][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:11,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:11,439][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:12,075][openai._base_client][INFO] - Retrying request to /chat/completions in 0.980236 seconds
+[2025-12-21 09:01:14,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:15,373][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:15,752][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493674 seconds
+[2025-12-21 09:01:15,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406426 seconds
+[2025-12-21 09:01:15,779][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466228 seconds
+[2025-12-21 09:01:15,808][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479543 seconds
+[2025-12-21 09:01:15,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391028 seconds
+[2025-12-21 09:01:15,831][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427245 seconds
+[2025-12-21 09:01:15,833][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466776 seconds
+[2025-12-21 09:01:15,834][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492788 seconds
+[2025-12-21 09:01:15,835][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472631 seconds
+[2025-12-21 09:01:15,836][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458378 seconds
+[2025-12-21 09:01:15,837][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400685 seconds
+[2025-12-21 09:01:15,838][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437556 seconds
+[2025-12-21 09:01:18,803][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:18,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:18,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:20,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:20,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:20,473][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:20,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432137 seconds
+[2025-12-21 09:01:20,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439478 seconds
+[2025-12-21 09:01:21,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:21,175][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939598 seconds
+[2025-12-21 09:01:21,224][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869742 seconds
+[2025-12-21 09:01:21,248][openai._base_client][INFO] - Retrying request to /chat/completions in 0.754146 seconds
+[2025-12-21 09:01:21,291][openai._base_client][INFO] - Retrying request to /chat/completions in 0.951066 seconds
+[2025-12-21 09:01:22,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:22,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:22,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:22,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:22,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:22,565][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:22,634][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:23,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:23,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:23,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:23,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:23,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:25,977][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:25,984][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:27,597][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:27,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:27,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:27,653][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:27,671][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:27,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:27,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:29,007][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:34,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424397 seconds
+[2025-12-21 09:01:36,166][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:36,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:39,040][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482556 seconds
+[2025-12-21 09:01:40,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:41,179][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499166 seconds
+[2025-12-21 09:01:41,312][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418943 seconds
+[2025-12-21 09:01:43,323][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:43,594][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:45,253][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:45,265][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:45,279][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:45,299][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:45,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:46,313][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403738 seconds
+[2025-12-21 09:01:46,315][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443877 seconds
+[2025-12-21 09:01:46,317][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406735 seconds
+[2025-12-21 09:01:46,317][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475957 seconds
+[2025-12-21 09:01:46,318][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469014 seconds
+[2025-12-21 09:01:46,322][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423581 seconds
+[2025-12-21 09:01:46,322][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383590 seconds
+[2025-12-21 09:01:46,323][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404667 seconds
+[2025-12-21 09:01:46,324][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494134 seconds
+[2025-12-21 09:01:46,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:46,680][openai._base_client][INFO] - Retrying request to /chat/completions in 0.820906 seconds
+[2025-12-21 09:01:46,738][openai._base_client][INFO] - Retrying request to /chat/completions in 0.819233 seconds
+[2025-12-21 09:01:46,860][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:46,882][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:46,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:47,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:48,637][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:49,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:49,242][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:49,256][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:49,264][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:49,395][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:49,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:50,040][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:50,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:51,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:51,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:51,166][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:51,719][openai._base_client][INFO] - Retrying request to /chat/completions in 0.855268 seconds
+[2025-12-21 09:01:51,726][openai._base_client][INFO] - Retrying request to /chat/completions in 0.852484 seconds
+[2025-12-21 09:01:51,730][openai._base_client][INFO] - Retrying request to /chat/completions in 0.840494 seconds
+[2025-12-21 09:01:51,748][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783376 seconds
+[2025-12-21 09:01:51,763][openai._base_client][INFO] - Retrying request to /chat/completions in 0.829933 seconds
+[2025-12-21 09:01:51,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.856590 seconds
+[2025-12-21 09:01:51,821][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921822 seconds
+[2025-12-21 09:01:53,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:53,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:53,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:53,976][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:54,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:54,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:54,291][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:01:59,309][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496075 seconds
+[2025-12-21 09:02:04,814][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899937 seconds
+[2025-12-21 09:02:17,587][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:18,125][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471305 seconds
+[2025-12-21 09:02:19,877][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:19,886][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:20,244][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:21,725][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:21,734][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:21,759][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:21,902][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:22,900][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:23,125][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381611 seconds
+[2025-12-21 09:02:23,126][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458620 seconds
+[2025-12-21 09:02:23,126][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414998 seconds
+[2025-12-21 09:02:23,127][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493379 seconds
+[2025-12-21 09:02:23,127][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463799 seconds
+[2025-12-21 09:02:23,128][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496611 seconds
+[2025-12-21 09:02:23,128][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477593 seconds
+[2025-12-21 09:02:23,129][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432457 seconds
+[2025-12-21 09:02:23,129][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472789 seconds
+[2025-12-21 09:02:23,130][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466584 seconds
+[2025-12-21 09:02:23,131][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489122 seconds
+[2025-12-21 09:02:23,522][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:23,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:23,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:23,599][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842742 seconds
+[2025-12-21 09:02:24,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:24,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:24,842][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:25,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:25,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:25,552][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:25,851][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:25,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:25,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:26,250][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:26,253][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:26,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:26,786][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:26,791][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:28,511][openai._base_client][INFO] - Retrying request to /chat/completions in 0.779894 seconds
+[2025-12-21 09:02:28,564][openai._base_client][INFO] - Retrying request to /chat/completions in 0.992433 seconds
+[2025-12-21 09:02:28,608][openai._base_client][INFO] - Retrying request to /chat/completions in 0.841580 seconds
+[2025-12-21 09:02:28,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901092 seconds
+[2025-12-21 09:02:28,625][openai._base_client][INFO] - Retrying request to /chat/completions in 0.758835 seconds
+[2025-12-21 09:02:28,627][openai._base_client][INFO] - Retrying request to /chat/completions in 0.998711 seconds
+[2025-12-21 09:02:30,716][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:30,728][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:30,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:31,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:31,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:31,533][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:02:36,555][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458606 seconds
+[2025-12-21 09:02:42,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871366 seconds
+[2025-12-21 09:02:57,921][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491806 seconds
+[2025-12-21 09:03:01,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:01,872][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488709 seconds
+[2025-12-21 09:03:03,810][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:03,950][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:04,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:04,404][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:04,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:04,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:04,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:04,750][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:05,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:05,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:05,421][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:05,606][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:05,716][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,085][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,283][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,725][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,737][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,872][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438224 seconds
+[2025-12-21 09:03:06,873][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444203 seconds
+[2025-12-21 09:03:06,874][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437886 seconds
+[2025-12-21 09:03:06,874][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420770 seconds
+[2025-12-21 09:03:06,875][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430258 seconds
+[2025-12-21 09:03:06,875][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418750 seconds
+[2025-12-21 09:03:06,922][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:06,928][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:07,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:07,509][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:07,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:08,137][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:08,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:08,882][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:08,888][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:09,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:09,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:09,403][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:03:11,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/09-06-15/.hydra/config.yaml b/outputs/2025-12-21/09-06-15/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a0e0ec7ad9ddf4236b55f4acf258e6615c38b0a6
--- /dev/null
+++ b/outputs/2025-12-21/09-06-15/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: fireworks/models/qwen3-235b-a22b-thinking-2507
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  fireworks/models/qwen3-235b-a22b-thinking-2507:
+    provider_name: openai
+    model_name: fireworks/models/qwen3-235b-a22b-thinking-2507
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/09-06-15/.hydra/hydra.yaml b/outputs/2025-12-21/09-06-15/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1b7c38e36bf9cbdd70fa3a9afda3d4ba185df3a6
--- /dev/null
+++ b/outputs/2025-12-21/09-06-15/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-06-15
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-06-15/.hydra/overrides.yaml b/outputs/2025-12-21/09-06-15/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-06-15/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-11-17/.hydra/config.yaml b/outputs/2025-12-21/09-11-17/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bdb3691d505e542cd59baf446c155a0d274edf2d
--- /dev/null
+++ b/outputs/2025-12-21/09-11-17/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-3-flash-preview
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-3-flash-preview:
+    provider_name: openai
+    model_name: gemini-3-flash-preview
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/09-11-17/.hydra/hydra.yaml b/outputs/2025-12-21/09-11-17/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..da7dfb954e253b0bb76d9fff1696348e07df23e4
--- /dev/null
+++ b/outputs/2025-12-21/09-11-17/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-11-17
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-11-17/.hydra/overrides.yaml b/outputs/2025-12-21/09-11-17/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-11-17/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-11-17/eval_api.log b/outputs/2025-12-21/09-11-17/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..f04bc60e462f0f92d93dd78d679833e78a093049
--- /dev/null
+++ b/outputs/2025-12-21/09-11-17/eval_api.log
@@ -0,0 +1,66 @@
+[2025-12-21 09:11:23,453][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428431 seconds
+[2025-12-21 09:11:23,454][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481088 seconds
+[2025-12-21 09:11:23,455][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407264 seconds
+[2025-12-21 09:11:23,455][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480485 seconds
+[2025-12-21 09:11:23,456][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376801 seconds
+[2025-12-21 09:11:23,456][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463087 seconds
+[2025-12-21 09:11:23,457][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418378 seconds
+[2025-12-21 09:11:23,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:23,726][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419418 seconds
+[2025-12-21 09:11:24,068][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:24,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:25,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:27,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:28,866][openai._base_client][INFO] - Retrying request to /chat/completions in 0.918882 seconds
+[2025-12-21 09:11:28,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.891364 seconds
+[2025-12-21 09:11:28,923][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864561 seconds
+[2025-12-21 09:11:28,939][openai._base_client][INFO] - Retrying request to /chat/completions in 0.850005 seconds
+[2025-12-21 09:11:28,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:29,192][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:29,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:30,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:30,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:33,773][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:34,341][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:34,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:34,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:35,565][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:45,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:11:50,936][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395835 seconds
+[2025-12-21 09:11:55,946][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491402 seconds
+[2025-12-21 09:11:55,947][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475317 seconds
+[2025-12-21 09:11:55,947][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396790 seconds
+[2025-12-21 09:12:02,479][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:03,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:07,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:07,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:10,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:15,431][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414104 seconds
+[2025-12-21 09:12:21,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:26,673][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:27,464][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:27,877][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:12:28,845][asyncio][ERROR] - Task exception was never retrieved
+future: <Task finished name='Task-93' coro=<ConcurrentLLM.generate() done, defined at /mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py:185> exception=ValueError('Content filtered or length exceeded')>
+Traceback (most recent call last):
+  File "/mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py", line 188, in generate
+    return await self.provider.generate(messages, **kwargs)
+  File "/mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py", line 49, in generate
+    raise ValueError("Content filtered or length exceeded")
+ValueError: Content filtered or length exceeded
+[2025-12-21 09:12:28,849][asyncio][ERROR] - Task exception was never retrieved
+future: <Task finished name='Task-92' coro=<ConcurrentLLM.generate() done, defined at /mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py:185> exception=ValueError('Content filtered or length exceeded')>
+Traceback (most recent call last):
+  File "/mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py", line 188, in generate
+    return await self.provider.generate(messages, **kwargs)
+  File "/mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py", line 49, in generate
+    raise ValueError("Content filtered or length exceeded")
+ValueError: Content filtered or length exceeded
+[2025-12-21 09:12:28,849][asyncio][ERROR] - Task exception was never retrieved
+future: <Task finished name='Task-90' coro=<ConcurrentLLM.generate() done, defined at /mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py:185> exception=ValueError('Content filtered or length exceeded')>
+Traceback (most recent call last):
+  File "/mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py", line 188, in generate
+    return await self.provider.generate(messages, **kwargs)
+  File "/mnt/general/wanghy/RAGEN/ragen/llm_agent/base_llm.py", line 49, in generate
+    raise ValueError("Content filtered or length exceeded")
+ValueError: Content filtered or length exceeded
diff --git a/outputs/2025-12-21/09-14-01/.hydra/config.yaml b/outputs/2025-12-21/09-14-01/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..11a11cc983c061f6f7dadc778d9f0017323f95e5
--- /dev/null
+++ b/outputs/2025-12-21/09-14-01/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 16
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 16
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/09-14-01/.hydra/hydra.yaml b/outputs/2025-12-21/09-14-01/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a3769fe672055e104474e5545c4c49b193966782
--- /dev/null
+++ b/outputs/2025-12-21/09-14-01/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-14-01
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-14-01/.hydra/overrides.yaml b/outputs/2025-12-21/09-14-01/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-14-01/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-14-01/eval_api.log b/outputs/2025-12-21/09-14-01/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..36fc1813b49871a312db0fa1fe60c6ad1d1dba51
--- /dev/null
+++ b/outputs/2025-12-21/09-14-01/eval_api.log
@@ -0,0 +1,124 @@
+[2025-12-21 09:14:07,524][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423758 seconds
+[2025-12-21 09:14:07,526][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459799 seconds
+[2025-12-21 09:14:07,526][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471574 seconds
+[2025-12-21 09:14:07,527][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467936 seconds
+[2025-12-21 09:14:07,527][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456045 seconds
+[2025-12-21 09:14:07,527][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376455 seconds
+[2025-12-21 09:14:07,527][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491551 seconds
+[2025-12-21 09:14:08,977][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:09,505][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:10,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:10,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:12,559][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:12,907][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920983 seconds
+[2025-12-21 09:14:12,952][openai._base_client][INFO] - Retrying request to /chat/completions in 0.784666 seconds
+[2025-12-21 09:14:12,986][openai._base_client][INFO] - Retrying request to /chat/completions in 0.923314 seconds
+[2025-12-21 09:14:12,989][openai._base_client][INFO] - Retrying request to /chat/completions in 0.861818 seconds
+[2025-12-21 09:14:12,997][openai._base_client][INFO] - Retrying request to /chat/completions in 0.995914 seconds
+[2025-12-21 09:14:13,000][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921900 seconds
+[2025-12-21 09:14:13,023][openai._base_client][INFO] - Retrying request to /chat/completions in 0.874127 seconds
+[2025-12-21 09:14:13,270][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:14,301][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:14,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:16,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:19,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:20,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:20,695][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:20,702][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:23,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:28,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:33,038][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420122 seconds
+[2025-12-21 09:14:38,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899798 seconds
+[2025-12-21 09:14:50,359][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:50,505][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475452 seconds
+[2025-12-21 09:14:54,737][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:55,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:55,238][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:55,360][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:55,465][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:55,506][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485898 seconds
+[2025-12-21 09:14:55,506][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388497 seconds
+[2025-12-21 09:14:55,507][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493651 seconds
+[2025-12-21 09:14:55,508][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389953 seconds
+[2025-12-21 09:14:55,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421240 seconds
+[2025-12-21 09:14:55,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447171 seconds
+[2025-12-21 09:14:55,510][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380228 seconds
+[2025-12-21 09:14:55,510][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401572 seconds
+[2025-12-21 09:14:56,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:56,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:58,702][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:59,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:59,650][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:14:59,982][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:00,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:00,996][openai._base_client][INFO] - Retrying request to /chat/completions in 0.778899 seconds
+[2025-12-21 09:15:01,002][openai._base_client][INFO] - Retrying request to /chat/completions in 0.961387 seconds
+[2025-12-21 09:15:01,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:04,957][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:05,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:05,102][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399058 seconds
+[2025-12-21 09:15:05,104][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383088 seconds
+[2025-12-21 09:15:05,105][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377228 seconds
+[2025-12-21 09:15:05,106][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437402 seconds
+[2025-12-21 09:15:10,109][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434344 seconds
+[2025-12-21 09:15:10,109][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377461 seconds
+[2025-12-21 09:15:10,485][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899092 seconds
+[2025-12-21 09:15:10,490][openai._base_client][INFO] - Retrying request to /chat/completions in 0.814934 seconds
+[2025-12-21 09:15:10,506][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807409 seconds
+[2025-12-21 09:15:10,546][openai._base_client][INFO] - Retrying request to /chat/completions in 0.851760 seconds
+[2025-12-21 09:15:10,949][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:14,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:14,847][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:15,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933472 seconds
+[2025-12-21 09:15:15,565][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:20,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:25,814][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468394 seconds
+[2025-12-21 09:15:29,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:30,825][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463454 seconds
+[2025-12-21 09:15:31,287][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899464 seconds
+[2025-12-21 09:15:35,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:42,201][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397827 seconds
+[2025-12-21 09:15:47,608][openai._base_client][INFO] - Retrying request to /chat/completions in 0.945020 seconds
+[2025-12-21 09:15:52,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:52,551][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383236 seconds
+[2025-12-21 09:15:56,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:15:57,553][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447229 seconds
+[2025-12-21 09:15:57,554][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444583 seconds
+[2025-12-21 09:15:57,554][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431193 seconds
+[2025-12-21 09:15:58,377][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:00,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:01,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:03,003][openai._base_client][INFO] - Retrying request to /chat/completions in 0.910123 seconds
+[2025-12-21 09:16:03,004][openai._base_client][INFO] - Retrying request to /chat/completions in 0.971744 seconds
+[2025-12-21 09:16:06,081][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:07,044][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:07,101][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458237 seconds
+[2025-12-21 09:16:07,103][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436004 seconds
+[2025-12-21 09:16:10,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:11,357][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:11,592][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:12,104][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466684 seconds
+[2025-12-21 09:16:12,104][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409287 seconds
+[2025-12-21 09:16:12,105][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424298 seconds
+[2025-12-21 09:16:14,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:15,394][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:15,587][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:15,639][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480618 seconds
+[2025-12-21 09:16:15,640][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465523 seconds
+[2025-12-21 09:16:15,642][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415584 seconds
+[2025-12-21 09:16:20,645][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470087 seconds
+[2025-12-21 09:16:21,061][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774101 seconds
+[2025-12-21 09:16:21,109][openai._base_client][INFO] - Retrying request to /chat/completions in 0.941332 seconds
+[2025-12-21 09:16:21,123][openai._base_client][INFO] - Retrying request to /chat/completions in 0.880977 seconds
+[2025-12-21 09:16:25,000][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:25,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:26,119][openai._base_client][INFO] - Retrying request to /chat/completions in 0.848411 seconds
+[2025-12-21 09:16:29,667][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:34,678][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408810 seconds
+[2025-12-21 09:16:39,780][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:39,826][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396247 seconds
+[2025-12-21 09:16:43,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:44,827][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468704 seconds
+[2025-12-21 09:16:47,620][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:16:47,651][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442374 seconds
+[2025-12-21 09:16:51,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/09-19-55/.hydra/config.yaml b/outputs/2025-12-21/09-19-55/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ef9e03828dc80841e3c81b38be15dbefd927251a
--- /dev/null
+++ b/outputs/2025-12-21/09-19-55/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 64
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 64
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 2048
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/09-19-55/.hydra/hydra.yaml b/outputs/2025-12-21/09-19-55/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bf10d5e9a07d08b4496183be889d01fc89b3dfd3
--- /dev/null
+++ b/outputs/2025-12-21/09-19-55/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-19-55
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-19-55/.hydra/overrides.yaml b/outputs/2025-12-21/09-19-55/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-19-55/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-19-55/eval_api.log b/outputs/2025-12-21/09-19-55/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..29a344bac66767f03d2e63e3945b28a24cc6072d
--- /dev/null
+++ b/outputs/2025-12-21/09-19-55/eval_api.log
@@ -0,0 +1,130 @@
+[2025-12-21 09:20:02,232][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475668 seconds
+[2025-12-21 09:20:02,233][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432879 seconds
+[2025-12-21 09:20:02,234][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395235 seconds
+[2025-12-21 09:20:02,235][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402898 seconds
+[2025-12-21 09:20:02,236][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381661 seconds
+[2025-12-21 09:20:02,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455038 seconds
+[2025-12-21 09:20:02,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461029 seconds
+[2025-12-21 09:20:02,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417799 seconds
+[2025-12-21 09:20:02,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480289 seconds
+[2025-12-21 09:20:02,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490984 seconds
+[2025-12-21 09:20:02,239][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488311 seconds
+[2025-12-21 09:20:02,239][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396960 seconds
+[2025-12-21 09:20:04,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:05,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:07,623][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908767 seconds
+[2025-12-21 09:20:07,632][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949494 seconds
+[2025-12-21 09:20:07,638][openai._base_client][INFO] - Retrying request to /chat/completions in 0.971473 seconds
+[2025-12-21 09:20:07,658][openai._base_client][INFO] - Retrying request to /chat/completions in 0.858431 seconds
+[2025-12-21 09:20:07,669][openai._base_client][INFO] - Retrying request to /chat/completions in 0.932979 seconds
+[2025-12-21 09:20:07,695][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755485 seconds
+[2025-12-21 09:20:07,699][openai._base_client][INFO] - Retrying request to /chat/completions in 0.826568 seconds
+[2025-12-21 09:20:07,712][openai._base_client][INFO] - Retrying request to /chat/completions in 0.934821 seconds
+[2025-12-21 09:20:07,720][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788562 seconds
+[2025-12-21 09:20:07,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.794073 seconds
+[2025-12-21 09:20:07,731][openai._base_client][INFO] - Retrying request to /chat/completions in 0.760326 seconds
+[2025-12-21 09:20:08,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:08,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:08,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:12,850][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:14,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:14,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:15,337][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:15,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:15,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:18,460][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446386 seconds
+[2025-12-21 09:20:18,499][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459006 seconds
+[2025-12-21 09:20:18,522][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380302 seconds
+[2025-12-21 09:20:18,537][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440813 seconds
+[2025-12-21 09:20:18,585][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425948 seconds
+[2025-12-21 09:20:18,608][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425586 seconds
+[2025-12-21 09:20:18,614][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419207 seconds
+[2025-12-21 09:20:18,654][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428057 seconds
+[2025-12-21 09:20:19,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:19,695][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:20,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:20,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:22,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:23,907][openai._base_client][INFO] - Retrying request to /chat/completions in 0.938769 seconds
+[2025-12-21 09:20:23,908][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797369 seconds
+[2025-12-21 09:20:23,962][openai._base_client][INFO] - Retrying request to /chat/completions in 0.881670 seconds
+[2025-12-21 09:20:23,982][openai._base_client][INFO] - Retrying request to /chat/completions in 0.793730 seconds
+[2025-12-21 09:20:24,038][openai._base_client][INFO] - Retrying request to /chat/completions in 0.879988 seconds
+[2025-12-21 09:20:24,038][openai._base_client][INFO] - Retrying request to /chat/completions in 0.953282 seconds
+[2025-12-21 09:20:24,084][openai._base_client][INFO] - Retrying request to /chat/completions in 0.889841 seconds
+[2025-12-21 09:20:24,858][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:25,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:26,608][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:28,740][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:29,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:30,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:30,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:31,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:31,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:32,804][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:33,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:34,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:34,511][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:35,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:35,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:36,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:36,945][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:37,261][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:37,280][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:38,059][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:38,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:39,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:39,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:40,593][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:41,508][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:42,116][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:43,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:44,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:44,260][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:44,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:44,940][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:45,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:46,437][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:46,744][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:48,388][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:49,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:49,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:50,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:20:55,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464906 seconds
+[2025-12-21 09:21:00,983][openai._base_client][INFO] - Retrying request to /chat/completions in 0.824079 seconds
+[2025-12-21 09:21:01,860][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:02,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:02,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:02,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:02,254][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:09,527][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:10,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:10,113][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:10,131][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:10,397][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:10,541][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:12,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:17,375][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477812 seconds
+[2025-12-21 09:21:22,386][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417456 seconds
+[2025-12-21 09:21:22,387][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424016 seconds
+[2025-12-21 09:21:22,857][openai._base_client][INFO] - Retrying request to /chat/completions in 0.994913 seconds
+[2025-12-21 09:21:27,818][openai._base_client][INFO] - Retrying request to /chat/completions in 0.868860 seconds
+[2025-12-21 09:21:31,710][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:32,068][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:37,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:41,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:21:46,314][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380925 seconds
+[2025-12-21 09:21:51,320][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428986 seconds
+[2025-12-21 09:21:51,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.961575 seconds
+[2025-12-21 09:21:56,757][openai._base_client][INFO] - Retrying request to /chat/completions in 0.924729 seconds
+[2025-12-21 09:22:00,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:22:07,694][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463322 seconds
+[2025-12-21 09:22:15,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:22:22,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:22:22,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:22:27,175][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485385 seconds
+[2025-12-21 09:22:41,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:22:42,137][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:22:47,148][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455870 seconds
+[2025-12-21 09:23:01,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:02,189][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/09-23-26/.hydra/config.yaml b/outputs/2025-12-21/09-23-26/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..272c2a6bfcbd3f8ffcd4e6537b19ce28f8d94559
--- /dev/null
+++ b/outputs/2025-12-21/09-23-26/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 64
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 64
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 4096
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/09-23-26/.hydra/hydra.yaml b/outputs/2025-12-21/09-23-26/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..66cd55aaf0db268487c7b2153acfe3a0287ea4c3
--- /dev/null
+++ b/outputs/2025-12-21/09-23-26/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-23-26
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-23-26/.hydra/overrides.yaml b/outputs/2025-12-21/09-23-26/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-23-26/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-23-26/eval_api.log b/outputs/2025-12-21/09-23-26/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..a53856a748aa03d3467751fb98643430465882c1
--- /dev/null
+++ b/outputs/2025-12-21/09-23-26/eval_api.log
@@ -0,0 +1,448 @@
+[2025-12-21 09:23:32,929][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458705 seconds
+[2025-12-21 09:23:32,930][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494792 seconds
+[2025-12-21 09:23:32,930][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484769 seconds
+[2025-12-21 09:23:32,931][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423300 seconds
+[2025-12-21 09:23:32,931][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396941 seconds
+[2025-12-21 09:23:33,476][openai._base_client][INFO] - Retrying request to /chat/completions in 0.789836 seconds
+[2025-12-21 09:23:33,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:34,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:34,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:34,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:34,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:35,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:35,755][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:39,134][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:39,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:40,341][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:40,419][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:40,491][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:40,566][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:40,652][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:40,664][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:41,182][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:41,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:41,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:44,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:45,241][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:45,455][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:45,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:46,433][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:47,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:47,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:48,214][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:48,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:49,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:49,566][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:51,060][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:51,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:51,424][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:51,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:52,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:54,631][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:54,903][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:55,566][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:56,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:56,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:57,271][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:57,741][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:58,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:58,427][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:58,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:58,805][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:23:59,738][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:00,632][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:00,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:01,383][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:01,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:01,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:03,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:03,669][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:04,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:05,448][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:06,644][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:07,221][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:07,963][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:07,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:08,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:08,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:10,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:13,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:15,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:20,188][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492521 seconds
+[2025-12-21 09:24:33,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:33,715][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497400 seconds
+[2025-12-21 09:24:37,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:37,251][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:37,743][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:37,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:37,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,207][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,364][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,665][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,869][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:38,935][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:39,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:40,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:41,288][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:41,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:41,519][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:41,602][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:41,977][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:42,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:42,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:42,922][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:43,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:43,435][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:43,622][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:44,068][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:44,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:44,223][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:44,530][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:44,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:45,294][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:45,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:45,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:45,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:46,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:46,143][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:46,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:46,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:47,187][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:47,350][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:47,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:47,643][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:48,124][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:48,316][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:48,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:49,112][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:49,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:49,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:50,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:50,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:50,182][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:50,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:50,478][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:50,884][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:52,013][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:52,313][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381102 seconds
+[2025-12-21 09:24:52,315][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433567 seconds
+[2025-12-21 09:24:52,318][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390586 seconds
+[2025-12-21 09:24:52,320][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377191 seconds
+[2025-12-21 09:24:52,322][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489915 seconds
+[2025-12-21 09:24:52,323][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392606 seconds
+[2025-12-21 09:24:52,325][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376789 seconds
+[2025-12-21 09:24:52,326][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497740 seconds
+[2025-12-21 09:24:52,327][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375252 seconds
+[2025-12-21 09:24:52,328][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451440 seconds
+[2025-12-21 09:24:52,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439424 seconds
+[2025-12-21 09:24:52,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494764 seconds
+[2025-12-21 09:24:52,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379231 seconds
+[2025-12-21 09:24:52,331][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383582 seconds
+[2025-12-21 09:24:52,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427218 seconds
+[2025-12-21 09:24:52,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394884 seconds
+[2025-12-21 09:24:55,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:56,583][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:56,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:56,763][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:56,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:57,190][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:57,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:57,485][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:57,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:57,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.768553 seconds
+[2025-12-21 09:24:57,703][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833675 seconds
+[2025-12-21 09:24:57,712][openai._base_client][INFO] - Retrying request to /chat/completions in 0.917243 seconds
+[2025-12-21 09:24:57,731][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:57,772][openai._base_client][INFO] - Retrying request to /chat/completions in 0.925837 seconds
+[2025-12-21 09:24:57,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.982010 seconds
+[2025-12-21 09:24:58,290][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:59,689][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:59,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:24:59,895][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:00,022][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:00,741][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:01,014][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:01,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:01,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:01,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:02,212][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:02,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:02,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:03,054][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:03,133][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:03,224][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:03,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:03,317][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:04,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:04,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:04,734][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:05,241][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:05,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:05,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:06,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:07,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:09,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:15,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:20,611][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399392 seconds
+[2025-12-21 09:25:26,018][openai._base_client][INFO] - Retrying request to /chat/completions in 0.946981 seconds
+[2025-12-21 09:25:39,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:40,219][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488399 seconds
+[2025-12-21 09:25:42,947][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:43,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:43,748][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:44,080][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:44,292][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:44,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:44,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:45,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:45,185][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:45,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:45,220][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384509 seconds
+[2025-12-21 09:25:45,221][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436878 seconds
+[2025-12-21 09:25:45,222][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492156 seconds
+[2025-12-21 09:25:45,222][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428855 seconds
+[2025-12-21 09:25:45,222][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380060 seconds
+[2025-12-21 09:25:45,711][openai._base_client][INFO] - Retrying request to /chat/completions in 0.932151 seconds
+[2025-12-21 09:25:45,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:46,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:47,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:47,674][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:47,954][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:48,283][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:48,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:48,723][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:48,957][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:49,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:49,786][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:50,595][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:50,654][openai._base_client][INFO] - Retrying request to /chat/completions in 0.874591 seconds
+[2025-12-21 09:25:50,660][openai._base_client][INFO] - Retrying request to /chat/completions in 0.981957 seconds
+[2025-12-21 09:25:50,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:50,717][openai._base_client][INFO] - Retrying request to /chat/completions in 0.758897 seconds
+[2025-12-21 09:25:50,938][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:51,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:52,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:53,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:55,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:55,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:25:56,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:26:15,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:26:20,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482004 seconds
+[2025-12-21 09:26:25,606][openai._base_client][INFO] - Retrying request to /chat/completions in 0.951439 seconds
+[2025-12-21 09:26:42,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:26:47,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399491 seconds
+[2025-12-21 09:26:52,878][openai._base_client][INFO] - Retrying request to /chat/completions in 0.932196 seconds
+[2025-12-21 09:26:57,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:26:58,028][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448464 seconds
+[2025-12-21 09:27:03,029][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429809 seconds
+[2025-12-21 09:27:03,030][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436975 seconds
+[2025-12-21 09:27:03,031][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446141 seconds
+[2025-12-21 09:27:03,031][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437333 seconds
+[2025-12-21 09:27:03,031][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482683 seconds
+[2025-12-21 09:27:03,031][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489955 seconds
+[2025-12-21 09:27:03,032][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455752 seconds
+[2025-12-21 09:27:03,032][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426325 seconds
+[2025-12-21 09:27:03,032][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463235 seconds
+[2025-12-21 09:27:03,032][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456440 seconds
+[2025-12-21 09:27:03,034][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461007 seconds
+[2025-12-21 09:27:03,034][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427161 seconds
+[2025-12-21 09:27:03,034][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439153 seconds
+[2025-12-21 09:27:03,034][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451894 seconds
+[2025-12-21 09:27:03,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449585 seconds
+[2025-12-21 09:27:03,480][openai._base_client][INFO] - Retrying request to /chat/completions in 0.759230 seconds
+[2025-12-21 09:27:08,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.841566 seconds
+[2025-12-21 09:27:08,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905025 seconds
+[2025-12-21 09:27:08,467][openai._base_client][INFO] - Retrying request to /chat/completions in 0.954949 seconds
+[2025-12-21 09:27:08,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863533 seconds
+[2025-12-21 09:27:08,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.990614 seconds
+[2025-12-21 09:27:08,476][openai._base_client][INFO] - Retrying request to /chat/completions in 0.986878 seconds
+[2025-12-21 09:27:08,479][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878608 seconds
+[2025-12-21 09:27:08,486][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921472 seconds
+[2025-12-21 09:27:08,488][openai._base_client][INFO] - Retrying request to /chat/completions in 0.760822 seconds
+[2025-12-21 09:27:08,489][openai._base_client][INFO] - Retrying request to /chat/completions in 0.769331 seconds
+[2025-12-21 09:27:08,490][openai._base_client][INFO] - Retrying request to /chat/completions in 0.950833 seconds
+[2025-12-21 09:27:08,498][openai._base_client][INFO] - Retrying request to /chat/completions in 0.855223 seconds
+[2025-12-21 09:27:08,498][openai._base_client][INFO] - Retrying request to /chat/completions in 0.945417 seconds
+[2025-12-21 09:27:08,516][openai._base_client][INFO] - Retrying request to /chat/completions in 0.886386 seconds
+[2025-12-21 09:27:08,523][openai._base_client][INFO] - Retrying request to /chat/completions in 0.965764 seconds
+[2025-12-21 09:27:14,253][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441206 seconds
+[2025-12-21 09:27:19,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443941 seconds
+[2025-12-21 09:27:19,264][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460260 seconds
+[2025-12-21 09:27:19,314][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473787 seconds
+[2025-12-21 09:27:19,339][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470828 seconds
+[2025-12-21 09:27:19,359][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403120 seconds
+[2025-12-21 09:27:19,364][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421961 seconds
+[2025-12-21 09:27:19,376][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407605 seconds
+[2025-12-21 09:27:19,408][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462599 seconds
+[2025-12-21 09:27:19,698][openai._base_client][INFO] - Retrying request to /chat/completions in 0.962571 seconds
+[2025-12-21 09:27:23,641][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:24,706][openai._base_client][INFO] - Retrying request to /chat/completions in 0.895539 seconds
+[2025-12-21 09:27:24,728][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774831 seconds
+[2025-12-21 09:27:24,766][openai._base_client][INFO] - Retrying request to /chat/completions in 0.770166 seconds
+[2025-12-21 09:27:24,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844777 seconds
+[2025-12-21 09:27:24,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.763258 seconds
+[2025-12-21 09:27:24,790][openai._base_client][INFO] - Retrying request to /chat/completions in 0.840823 seconds
+[2025-12-21 09:27:24,873][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860012 seconds
+[2025-12-21 09:27:28,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:35,745][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454433 seconds
+[2025-12-21 09:27:39,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:40,784][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413456 seconds
+[2025-12-21 09:27:40,785][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432816 seconds
+[2025-12-21 09:27:40,785][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406306 seconds
+[2025-12-21 09:27:40,785][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391065 seconds
+[2025-12-21 09:27:40,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404038 seconds
+[2025-12-21 09:27:40,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401986 seconds
+[2025-12-21 09:27:40,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423055 seconds
+[2025-12-21 09:27:40,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406107 seconds
+[2025-12-21 09:27:40,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391294 seconds
+[2025-12-21 09:27:40,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377233 seconds
+[2025-12-21 09:27:40,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456808 seconds
+[2025-12-21 09:27:40,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446099 seconds
+[2025-12-21 09:27:40,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451830 seconds
+[2025-12-21 09:27:40,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446102 seconds
+[2025-12-21 09:27:40,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431878 seconds
+[2025-12-21 09:27:42,696][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:44,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:45,359][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:46,167][openai._base_client][INFO] - Retrying request to /chat/completions in 0.834475 seconds
+[2025-12-21 09:27:46,178][openai._base_client][INFO] - Retrying request to /chat/completions in 0.902990 seconds
+[2025-12-21 09:27:46,180][openai._base_client][INFO] - Retrying request to /chat/completions in 0.835472 seconds
+[2025-12-21 09:27:46,190][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955706 seconds
+[2025-12-21 09:27:46,191][openai._base_client][INFO] - Retrying request to /chat/completions in 0.866868 seconds
+[2025-12-21 09:27:46,193][openai._base_client][INFO] - Retrying request to /chat/completions in 0.958844 seconds
+[2025-12-21 09:27:46,201][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916160 seconds
+[2025-12-21 09:27:46,211][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869403 seconds
+[2025-12-21 09:27:46,220][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949091 seconds
+[2025-12-21 09:27:46,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.834764 seconds
+[2025-12-21 09:27:46,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.895257 seconds
+[2025-12-21 09:27:46,241][openai._base_client][INFO] - Retrying request to /chat/completions in 0.885029 seconds
+[2025-12-21 09:27:46,931][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:47,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:49,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:49,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:50,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:50,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:50,651][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:51,019][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:51,149][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:51,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:51,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:51,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:27:57,180][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426023 seconds
+[2025-12-21 09:27:59,873][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:00,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:00,159][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:00,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:00,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:01,489][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:02,612][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797370 seconds
+[2025-12-21 09:28:06,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:07,034][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397701 seconds
+[2025-12-21 09:28:12,039][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494154 seconds
+[2025-12-21 09:28:12,040][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427569 seconds
+[2025-12-21 09:28:12,041][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490461 seconds
+[2025-12-21 09:28:12,041][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401918 seconds
+[2025-12-21 09:28:12,041][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409225 seconds
+[2025-12-21 09:28:12,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393628 seconds
+[2025-12-21 09:28:12,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387837 seconds
+[2025-12-21 09:28:12,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448653 seconds
+[2025-12-21 09:28:12,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390900 seconds
+[2025-12-21 09:28:12,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446567 seconds
+[2025-12-21 09:28:12,043][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379176 seconds
+[2025-12-21 09:28:12,043][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419649 seconds
+[2025-12-21 09:28:12,043][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483757 seconds
+[2025-12-21 09:28:12,435][openai._base_client][INFO] - Retrying request to /chat/completions in 0.885913 seconds
+[2025-12-21 09:28:17,429][openai._base_client][INFO] - Retrying request to /chat/completions in 0.751148 seconds
+[2025-12-21 09:28:17,432][openai._base_client][INFO] - Retrying request to /chat/completions in 0.770374 seconds
+[2025-12-21 09:28:17,435][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816740 seconds
+[2025-12-21 09:28:17,437][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767661 seconds
+[2025-12-21 09:28:17,445][openai._base_client][INFO] - Retrying request to /chat/completions in 0.993458 seconds
+[2025-12-21 09:28:17,453][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833944 seconds
+[2025-12-21 09:28:17,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.826601 seconds
+[2025-12-21 09:28:17,470][openai._base_client][INFO] - Retrying request to /chat/completions in 0.928424 seconds
+[2025-12-21 09:28:17,493][openai._base_client][INFO] - Retrying request to /chat/completions in 0.802728 seconds
+[2025-12-21 09:28:17,493][openai._base_client][INFO] - Retrying request to /chat/completions in 0.805865 seconds
+[2025-12-21 09:28:17,529][openai._base_client][INFO] - Retrying request to /chat/completions in 0.990451 seconds
+[2025-12-21 09:28:17,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.969602 seconds
+[2025-12-21 09:28:17,536][openai._base_client][INFO] - Retrying request to /chat/completions in 0.900609 seconds
+[2025-12-21 09:28:31,303][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:31,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:31,477][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:32,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:33,576][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404391 seconds
+[2025-12-21 09:28:33,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382638 seconds
+[2025-12-21 09:28:33,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412469 seconds
+[2025-12-21 09:28:33,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461459 seconds
+[2025-12-21 09:28:33,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394651 seconds
+[2025-12-21 09:28:33,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417424 seconds
+[2025-12-21 09:28:33,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431000 seconds
+[2025-12-21 09:28:33,579][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432827 seconds
+[2025-12-21 09:28:33,579][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476703 seconds
+[2025-12-21 09:28:34,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:36,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:36,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:36,909][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:37,489][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:37,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:37,732][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:38,999][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967643 seconds
+[2025-12-21 09:28:39,014][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860220 seconds
+[2025-12-21 09:28:39,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.885488 seconds
+[2025-12-21 09:28:42,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:43,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:44,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:44,722][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461475 seconds
+[2025-12-21 09:28:44,730][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382318 seconds
+[2025-12-21 09:28:44,744][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440258 seconds
+[2025-12-21 09:28:47,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:48,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:48,636][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:48,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:49,739][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440706 seconds
+[2025-12-21 09:28:49,740][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459158 seconds
+[2025-12-21 09:28:49,743][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492773 seconds
+[2025-12-21 09:28:49,748][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396317 seconds
+[2025-12-21 09:28:49,764][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:50,190][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807492 seconds
+[2025-12-21 09:28:52,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:52,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:53,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:53,953][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:55,124][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:55,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453715 seconds
+[2025-12-21 09:28:55,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385014 seconds
+[2025-12-21 09:28:55,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477838 seconds
+[2025-12-21 09:28:55,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388527 seconds
+[2025-12-21 09:28:55,241][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480443 seconds
+[2025-12-21 09:28:58,403][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:58,451][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:58,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:58,595][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:58,596][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:58,672][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:28:58,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:00,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455581 seconds
+[2025-12-21 09:29:02,979][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:03,099][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485243 seconds
+[2025-12-21 09:29:03,100][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480189 seconds
+[2025-12-21 09:29:03,101][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376118 seconds
+[2025-12-21 09:29:03,102][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382939 seconds
+[2025-12-21 09:29:03,103][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416352 seconds
+[2025-12-21 09:29:03,480][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812349 seconds
+[2025-12-21 09:29:03,487][openai._base_client][INFO] - Retrying request to /chat/completions in 0.965780 seconds
+[2025-12-21 09:29:06,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:06,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:07,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:08,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.909316 seconds
+[2025-12-21 09:29:08,913][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:11,876][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:11,922][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484412 seconds
+[2025-12-21 09:29:11,923][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421127 seconds
+[2025-12-21 09:29:12,346][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783831 seconds
+[2025-12-21 09:29:12,409][openai._base_client][INFO] - Retrying request to /chat/completions in 0.868666 seconds
+[2025-12-21 09:29:15,639][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:29:27,915][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/09-35-01/.hydra/config.yaml b/outputs/2025-12-21/09-35-01/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..449aff9b3778849d89e77d4464f2ceb41617dcaa
--- /dev/null
+++ b/outputs/2025-12-21/09-35-01/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 300
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 4096
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/09-35-01/.hydra/hydra.yaml b/outputs/2025-12-21/09-35-01/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..2b2586f4dba9c7ca95ec972e5a12d7a379157bfd
--- /dev/null
+++ b/outputs/2025-12-21/09-35-01/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-35-01
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-35-01/.hydra/overrides.yaml b/outputs/2025-12-21/09-35-01/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-35-01/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-35-01/eval_api.log b/outputs/2025-12-21/09-35-01/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..5931facf94f4ea0565d9f50d0bf5c4b3ebe72bef
--- /dev/null
+++ b/outputs/2025-12-21/09-35-01/eval_api.log
@@ -0,0 +1,17 @@
+[2025-12-21 09:35:08,375][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458412 seconds
+[2025-12-21 09:35:13,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.946658 seconds
+[2025-12-21 09:35:36,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:35:36,539][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399683 seconds
+[2025-12-21 09:35:41,944][openai._base_client][INFO] - Retrying request to /chat/completions in 0.971540 seconds
+[2025-12-21 09:35:51,855][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:35:51,867][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388661 seconds
+[2025-12-21 09:36:04,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:36:04,718][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436150 seconds
+[2025-12-21 09:36:35,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:36:40,217][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402573 seconds
+[2025-12-21 09:37:11,521][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:37:16,532][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444043 seconds
+[2025-12-21 09:37:46,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:37:51,593][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392293 seconds
+[2025-12-21 09:38:22,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:38:27,873][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425745 seconds
diff --git a/outputs/2025-12-21/09-38-58/.hydra/config.yaml b/outputs/2025-12-21/09-38-58/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ff289d30b3ac47bc1156b8a053a950ce18c8851a
--- /dev/null
+++ b/outputs/2025-12-21/09-38-58/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 300
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/09-38-58/.hydra/hydra.yaml b/outputs/2025-12-21/09-38-58/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..1de5dd8b029ba57e594aca78a09221370a647408
--- /dev/null
+++ b/outputs/2025-12-21/09-38-58/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-38-58
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-38-58/.hydra/overrides.yaml b/outputs/2025-12-21/09-38-58/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-38-58/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-38-58/eval_api.log b/outputs/2025-12-21/09-38-58/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..849b1490d51506c88205a1ed576790a6de5dcc08
--- /dev/null
+++ b/outputs/2025-12-21/09-38-58/eval_api.log
@@ -0,0 +1,117 @@
+[2025-12-21 09:39:11,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:39:11,590][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498221 seconds
+[2025-12-21 09:39:20,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:39:20,646][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431428 seconds
+[2025-12-21 09:39:26,084][openai._base_client][INFO] - Retrying request to /chat/completions in 0.750108 seconds
+[2025-12-21 09:39:41,852][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407842 seconds
+[2025-12-21 09:39:54,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:39:54,318][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486661 seconds
+[2025-12-21 09:39:59,813][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797331 seconds
+[2025-12-21 09:40:15,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490150 seconds
+[2025-12-21 09:40:21,128][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933661 seconds
+[2025-12-21 09:40:37,083][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399590 seconds
+[2025-12-21 09:40:42,997][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:40:43,032][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472312 seconds
+[2025-12-21 09:40:48,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:40:48,919][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417444 seconds
+[2025-12-21 09:40:54,344][openai._base_client][INFO] - Retrying request to /chat/completions in 0.911699 seconds
+[2025-12-21 09:41:10,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421365 seconds
+[2025-12-21 09:41:16,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:41:16,586][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408452 seconds
+[2025-12-21 09:41:22,002][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944636 seconds
+[2025-12-21 09:41:37,971][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443674 seconds
+[2025-12-21 09:41:45,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:41:45,154][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463111 seconds
+[2025-12-21 09:41:50,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.855337 seconds
+[2025-12-21 09:42:06,503][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391662 seconds
+[2025-12-21 09:42:27,196][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:42:27,221][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 7162).
+[2025-12-21 09:42:27,242][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483890 seconds
+[2025-12-21 09:42:34,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:42:34,143][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6647).
+[2025-12-21 09:42:34,157][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466779 seconds
+[2025-12-21 09:42:40,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:42:40,696][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6676).
+[2025-12-21 09:42:40,713][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413958 seconds
+[2025-12-21 09:42:47,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:42:48,000][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6810).
+[2025-12-21 09:42:48,017][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487181 seconds
+[2025-12-21 09:42:54,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:42:55,012][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6829).
+[2025-12-21 09:42:55,028][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415029 seconds
+[2025-12-21 09:43:01,684][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:43:01,720][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6877).
+[2025-12-21 09:43:01,736][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473990 seconds
+[2025-12-21 09:43:08,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:43:08,085][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 4770).
+[2025-12-21 09:43:08,097][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443073 seconds
+[2025-12-21 09:43:14,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:43:14,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473509 seconds
+[2025-12-21 09:43:20,618][openai._base_client][INFO] - Retrying request to /chat/completions in 0.828059 seconds
+[2025-12-21 09:43:41,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:43:41,346][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6847).
+[2025-12-21 09:43:41,365][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415223 seconds
+[2025-12-21 09:43:48,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:43:48,705][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6858).
+[2025-12-21 09:43:48,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405821 seconds
+[2025-12-21 09:43:54,409][openai._base_client][INFO] - Retrying request to /chat/completions in 0.809677 seconds
+[2025-12-21 09:44:01,295][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433622 seconds
+[2025-12-21 09:44:06,737][openai._base_client][INFO] - Retrying request to /chat/completions in 0.834175 seconds
+[2025-12-21 09:44:25,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:44:25,871][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6903).
+[2025-12-21 09:44:25,896][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430662 seconds
+[2025-12-21 09:44:34,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:44:34,773][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6587).
+[2025-12-21 09:44:34,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396953 seconds
+[2025-12-21 09:44:43,268][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:44:43,296][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 4816).
+[2025-12-21 09:44:43,310][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391169 seconds
+[2025-12-21 09:44:44,225][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755211 seconds
+[2025-12-21 09:45:04,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:45:04,606][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6714).
+[2025-12-21 09:45:04,622][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484243 seconds
+[2025-12-21 09:45:11,414][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:45:11,438][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 5853).
+[2025-12-21 09:45:11,451][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379683 seconds
+[2025-12-21 09:45:28,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:45:28,374][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 7115).
+[2025-12-21 09:45:28,394][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389686 seconds
+[2025-12-21 09:45:42,044][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:45:42,079][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 5139).
+[2025-12-21 09:45:42,092][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499970 seconds
+[2025-12-21 09:46:01,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:46:01,534][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6997).
+[2025-12-21 09:46:01,549][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393224 seconds
+[2025-12-21 09:46:10,416][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:46:10,438][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 5301).
+[2025-12-21 09:46:10,472][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482374 seconds
+[2025-12-21 09:46:16,598][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:46:16,631][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474281 seconds
+[2025-12-21 09:46:17,621][openai._base_client][INFO] - Retrying request to /chat/completions in 0.883785 seconds
+[2025-12-21 09:46:29,027][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481792 seconds
+[2025-12-21 09:46:34,517][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979346 seconds
+[2025-12-21 09:46:47,087][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:46:47,116][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400456 seconds
+[2025-12-21 09:46:52,524][openai._base_client][INFO] - Retrying request to /chat/completions in 0.974026 seconds
+[2025-12-21 09:47:07,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:47:07,260][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 5596).
+[2025-12-21 09:47:07,274][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407225 seconds
+[2025-12-21 09:48:40,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:48:45,975][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424250 seconds
+[2025-12-21 09:49:00,140][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:49:00,169][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444292 seconds
+[2025-12-21 09:49:21,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:49:21,061][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379879 seconds
+[2025-12-21 09:49:28,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:49:28,983][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 5489).
+[2025-12-21 09:49:28,997][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378431 seconds
+[2025-12-21 09:49:34,383][openai._base_client][INFO] - Retrying request to /chat/completions in 0.866210 seconds
+[2025-12-21 09:49:41,838][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:49:41,879][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432670 seconds
+[2025-12-21 09:49:47,319][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869630 seconds
+[2025-12-21 09:49:55,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:49:55,611][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441269 seconds
+[2025-12-21 09:50:01,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.768853 seconds
+[2025-12-21 09:50:07,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:50:07,957][root][WARNING] - Truncated history for env 0 to keep prompt within 7200 tokens (current 6674).
+[2025-12-21 09:50:07,974][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386858 seconds
diff --git a/outputs/2025-12-21/09-51-06/.hydra/config.yaml b/outputs/2025-12-21/09-51-06/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..163fd6384b97cdaafe90e689cfac621ccafcbdf6
--- /dev/null
+++ b/outputs/2025-12-21/09-51-06/.hydra/config.yaml
@@ -0,0 +1,920 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 64000
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 300
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
diff --git a/outputs/2025-12-21/09-51-06/.hydra/hydra.yaml b/outputs/2025-12-21/09-51-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..61107bfab74b495fe7c7727f8ed508bd7c476248
--- /dev/null
+++ b/outputs/2025-12-21/09-51-06/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/09-51-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/09-51-06/.hydra/overrides.yaml b/outputs/2025-12-21/09-51-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/09-51-06/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/09-51-06/eval_api.log b/outputs/2025-12-21/09-51-06/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..ec62d5021d8f9c6a5718b1e59c22264f39508108
--- /dev/null
+++ b/outputs/2025-12-21/09-51-06/eval_api.log
@@ -0,0 +1,251 @@
+[2025-12-21 09:51:19,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:51:19,259][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458445 seconds
+[2025-12-21 09:51:28,194][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:51:28,206][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470919 seconds
+[2025-12-21 09:51:41,448][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:51:41,462][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471261 seconds
+[2025-12-21 09:51:47,428][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:51:47,444][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442771 seconds
+[2025-12-21 09:51:55,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:51:55,895][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461484 seconds
+[2025-12-21 09:52:05,092][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:52:05,115][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426794 seconds
+[2025-12-21 09:52:13,337][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:52:13,365][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432980 seconds
+[2025-12-21 09:52:21,714][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:52:21,742][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379941 seconds
+[2025-12-21 09:52:30,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:52:30,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454259 seconds
+[2025-12-21 09:52:37,651][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:52:37,689][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389561 seconds
+[2025-12-21 09:52:45,076][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:52:45,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393990 seconds
+[2025-12-21 09:52:52,004][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:52:52,044][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420928 seconds
+[2025-12-21 09:53:11,588][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:53:11,634][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435929 seconds
+[2025-12-21 09:53:18,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:53:18,503][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376671 seconds
+[2025-12-21 09:53:26,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:53:26,787][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395572 seconds
+[2025-12-21 09:53:42,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:53:42,366][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447784 seconds
+[2025-12-21 09:53:50,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:53:50,662][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482226 seconds
+[2025-12-21 09:54:00,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:54:00,908][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476984 seconds
+[2025-12-21 09:54:01,550][openai._base_client][INFO] - Retrying request to /chat/completions in 0.934005 seconds
+[2025-12-21 09:54:08,243][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:54:08,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422204 seconds
+[2025-12-21 09:54:08,886][openai._base_client][INFO] - Retrying request to /chat/completions in 0.994106 seconds
+[2025-12-21 09:54:15,201][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493208 seconds
+[2025-12-21 09:54:21,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:54:21,549][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476922 seconds
+[2025-12-21 09:54:29,633][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:54:29,709][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387497 seconds
+[2025-12-21 09:54:30,252][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933914 seconds
+[2025-12-21 09:54:38,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:54:38,540][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414123 seconds
+[2025-12-21 09:54:39,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.993401 seconds
+[2025-12-21 09:54:50,286][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463455 seconds
+[2025-12-21 09:55:01,160][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:55:01,248][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387065 seconds
+[2025-12-21 09:55:08,243][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:55:08,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409273 seconds
+[2025-12-21 09:55:17,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:55:17,688][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493643 seconds
+[2025-12-21 09:55:24,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:55:25,054][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379996 seconds
+[2025-12-21 09:55:32,056][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:55:32,173][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406281 seconds
+[2025-12-21 09:55:32,736][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904648 seconds
+[2025-12-21 09:55:39,940][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:55:40,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437361 seconds
+[2025-12-21 09:55:45,471][openai._base_client][INFO] - Retrying request to /chat/completions in 0.975070 seconds
+[2025-12-21 09:55:55,140][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:55:55,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484819 seconds
+[2025-12-21 09:56:00,722][openai._base_client][INFO] - Retrying request to /chat/completions in 0.935471 seconds
+[2025-12-21 09:56:11,687][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:56:11,777][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466472 seconds
+[2025-12-21 09:56:17,811][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:56:17,903][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480557 seconds
+[2025-12-21 09:56:25,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:56:25,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497981 seconds
+[2025-12-21 09:56:32,419][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:56:32,513][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449268 seconds
+[2025-12-21 09:56:44,909][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:56:45,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454414 seconds
+[2025-12-21 09:56:45,643][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870614 seconds
+[2025-12-21 09:56:53,657][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:56:53,760][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432142 seconds
+[2025-12-21 09:57:12,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:57:12,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410561 seconds
+[2025-12-21 09:57:17,534][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781510 seconds
+[2025-12-21 09:57:32,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:57:33,066][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452802 seconds
+[2025-12-21 09:57:43,001][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:57:43,137][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408897 seconds
+[2025-12-21 09:57:51,825][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:57:51,938][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476704 seconds
+[2025-12-21 09:58:06,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:58:06,454][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388704 seconds
+[2025-12-21 09:58:15,066][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:58:15,188][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420927 seconds
+[2025-12-21 09:58:23,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:58:23,674][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378771 seconds
+[2025-12-21 09:58:33,848][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:58:34,005][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456069 seconds
+[2025-12-21 09:58:44,529][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:58:44,664][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404120 seconds
+[2025-12-21 09:58:52,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:58:53,038][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447109 seconds
+[2025-12-21 09:58:58,494][openai._base_client][INFO] - Retrying request to /chat/completions in 0.759045 seconds
+[2025-12-21 09:59:14,295][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415154 seconds
+[2025-12-21 09:59:19,720][openai._base_client][INFO] - Retrying request to /chat/completions in 0.953460 seconds
+[2025-12-21 09:59:35,712][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475830 seconds
+[2025-12-21 09:59:43,368][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 09:59:43,552][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488879 seconds
+[2025-12-21 09:59:49,049][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920177 seconds
+[2025-12-21 10:00:13,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:00:13,179][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482228 seconds
+[2025-12-21 10:00:18,816][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:00:18,973][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416419 seconds
+[2025-12-21 10:00:24,398][openai._base_client][INFO] - Retrying request to /chat/completions in 0.834784 seconds
+[2025-12-21 10:00:33,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:00:33,913][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388761 seconds
+[2025-12-21 10:00:39,310][openai._base_client][INFO] - Retrying request to /chat/completions in 0.896599 seconds
+[2025-12-21 10:00:48,212][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:00:48,410][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486092 seconds
+[2025-12-21 10:00:56,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:00:56,845][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381071 seconds
+[2025-12-21 10:01:03,198][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:01:03,358][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391693 seconds
+[2025-12-21 10:01:11,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:01:11,963][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492237 seconds
+[2025-12-21 10:01:19,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:01:20,044][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488139 seconds
+[2025-12-21 10:01:27,562][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:01:27,789][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491307 seconds
+[2025-12-21 10:01:36,035][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:01:36,225][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387867 seconds
+[2025-12-21 10:01:45,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:01:46,187][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403928 seconds
+[2025-12-21 10:01:51,600][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985951 seconds
+[2025-12-21 10:02:07,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405111 seconds
+[2025-12-21 10:02:13,039][openai._base_client][INFO] - Retrying request to /chat/completions in 0.982400 seconds
+[2025-12-21 10:02:26,032][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:02:26,311][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457664 seconds
+[2025-12-21 10:02:33,047][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:02:33,221][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460037 seconds
+[2025-12-21 10:02:38,690][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774831 seconds
+[2025-12-21 10:02:47,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:02:47,677][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430614 seconds
+[2025-12-21 10:02:55,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:02:55,394][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434698 seconds
+[2025-12-21 10:03:03,024][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:03:03,250][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398996 seconds
+[2025-12-21 10:03:12,882][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:03:13,121][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408968 seconds
+[2025-12-21 10:03:22,723][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:03:22,953][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409830 seconds
+[2025-12-21 10:03:29,280][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:03:29,486][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405496 seconds
+[2025-12-21 10:03:34,900][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972057 seconds
+[2025-12-21 10:03:50,914][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499139 seconds
+[2025-12-21 10:03:57,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:03:57,774][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449255 seconds
+[2025-12-21 10:04:03,231][openai._base_client][INFO] - Retrying request to /chat/completions in 0.858854 seconds
+[2025-12-21 10:04:19,132][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407206 seconds
+[2025-12-21 10:04:24,549][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899697 seconds
+[2025-12-21 10:04:36,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:04:36,838][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383989 seconds
+[2025-12-21 10:04:48,282][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:04:48,566][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494426 seconds
+[2025-12-21 10:04:54,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:04:54,668][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397315 seconds
+[2025-12-21 10:05:07,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:05:08,078][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380333 seconds
+[2025-12-21 10:05:16,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:05:17,095][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474107 seconds
+[2025-12-21 10:05:31,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:05:31,658][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63847).
+[2025-12-21 10:05:31,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389327 seconds
+[2025-12-21 10:07:09,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:07:15,013][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478031 seconds
+[2025-12-21 10:08:52,806][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:08:58,117][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450148 seconds
+[2025-12-21 10:10:34,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:10:39,373][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424841 seconds
+[2025-12-21 10:10:47,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:10:47,937][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63408).
+[2025-12-21 10:10:48,070][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496255 seconds
+[2025-12-21 10:10:59,163][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:10:59,345][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63709).
+[2025-12-21 10:10:59,532][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399288 seconds
+[2025-12-21 10:11:09,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:11:09,346][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63749).
+[2025-12-21 10:11:09,482][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418732 seconds
+[2025-12-21 10:11:19,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:11:19,444][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63578).
+[2025-12-21 10:11:19,578][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489078 seconds
+[2025-12-21 10:11:29,558][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:11:29,726][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63563).
+[2025-12-21 10:11:29,854][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456990 seconds
+[2025-12-21 10:11:40,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:11:40,823][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63685).
+[2025-12-21 10:11:40,972][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393813 seconds
+[2025-12-21 10:11:51,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:11:51,805][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63655).
+[2025-12-21 10:11:51,931][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451145 seconds
+[2025-12-21 10:12:04,063][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:12:04,285][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63992).
+[2025-12-21 10:12:04,417][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470850 seconds
+[2025-12-21 10:12:13,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:12:13,479][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63287).
+[2025-12-21 10:12:13,603][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475179 seconds
+[2025-12-21 10:12:25,123][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:12:25,381][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 61758).
+[2025-12-21 10:12:25,495][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461899 seconds
+[2025-12-21 10:12:36,813][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:12:37,124][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439757 seconds
+[2025-12-21 10:12:46,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:12:47,158][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481428 seconds
+[2025-12-21 10:12:52,649][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774632 seconds
+[2025-12-21 10:13:12,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:13:13,242][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63374).
+[2025-12-21 10:13:13,372][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386459 seconds
+[2025-12-21 10:13:23,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:13:23,821][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 62709).
+[2025-12-21 10:13:23,973][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470977 seconds
+[2025-12-21 10:13:29,453][openai._base_client][INFO] - Retrying request to /chat/completions in 0.789298 seconds
+[2025-12-21 10:13:38,045][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:13:38,266][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481385 seconds
+[2025-12-21 10:13:47,672][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:13:47,827][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63491).
+[2025-12-21 10:13:47,955][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471657 seconds
+[2025-12-21 10:13:58,167][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:13:58,360][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63555).
+[2025-12-21 10:13:58,489][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407000 seconds
+[2025-12-21 10:14:11,494][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 10:14:11,690][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63983).
+[2025-12-21 10:14:11,853][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476169 seconds
+[2025-12-21 10:14:15,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:14:20,078][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422746 seconds
+[2025-12-21 10:14:22,073][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:14:27,105][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451933 seconds
+[2025-12-21 10:14:29,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:14:34,138][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434184 seconds
+[2025-12-21 10:14:39,581][openai._base_client][INFO] - Retrying request to /chat/completions in 0.885823 seconds
+[2025-12-21 10:14:42,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:14:47,060][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474222 seconds
+[2025-12-21 10:14:49,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:14:54,419][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408779 seconds
+[2025-12-21 10:14:56,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:15:01,393][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398547 seconds
+[2025-12-21 10:15:03,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:15:08,447][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423065 seconds
+[2025-12-21 10:15:10,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:15:15,451][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427477 seconds
+[2025-12-21 10:15:17,489][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
+[2025-12-21 10:15:22,517][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450296 seconds
+[2025-12-21 10:15:24,560][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 402 Payment Required"
diff --git a/outputs/2025-12-21/11-05-59/.hydra/config.yaml b/outputs/2025-12-21/11-05-59/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..b99f11c3a4bcf6e47b83c99c798803c2cb8617ae
--- /dev/null
+++ b/outputs/2025-12-21/11-05-59/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 64000
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 300
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 4096
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/11-05-59/.hydra/hydra.yaml b/outputs/2025-12-21/11-05-59/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..5a01ee61d978d3ea50eda7e80d046c1c79f26992
--- /dev/null
+++ b/outputs/2025-12-21/11-05-59/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/11-05-59
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/11-05-59/.hydra/overrides.yaml b/outputs/2025-12-21/11-05-59/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/11-05-59/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/11-05-59/eval_api.log b/outputs/2025-12-21/11-05-59/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..efdcd3626b025f66bf628551a56438a03df55c48
--- /dev/null
+++ b/outputs/2025-12-21/11-05-59/eval_api.log
@@ -0,0 +1,132 @@
+[2025-12-21 11:06:01,872][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491323 seconds
+[2025-12-21 11:06:06,428][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464464 seconds
+[2025-12-21 11:06:06,429][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490795 seconds
+[2025-12-21 11:06:06,701][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426827 seconds
+[2025-12-21 11:06:07,366][openai._base_client][INFO] - Retrying request to /chat/completions in 0.997493 seconds
+[2025-12-21 11:06:07,537][openai._base_client][INFO] - Retrying request to /chat/completions in 0.917339 seconds
+[2025-12-21 11:06:13,842][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421775 seconds
+[2025-12-21 11:06:14,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:14,651][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:15,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:16,291][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:17,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:17,915][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:18,679][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:19,268][openai._base_client][INFO] - Retrying request to /chat/completions in 0.793220 seconds
+[2025-12-21 11:06:22,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:27,189][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:29,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:29,425][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:29,640][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:30,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:30,113][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:31,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:32,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:34,831][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:35,428][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:40,897][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:41,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:42,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:42,643][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:43,133][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:43,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:44,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:46,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:46,736][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:49,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:06:56,572][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:01,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487332 seconds
+[2025-12-21 11:07:06,598][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421268 seconds
+[2025-12-21 11:07:06,600][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402390 seconds
+[2025-12-21 11:07:06,600][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441221 seconds
+[2025-12-21 11:07:06,600][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476082 seconds
+[2025-12-21 11:07:06,601][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467352 seconds
+[2025-12-21 11:07:06,601][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460949 seconds
+[2025-12-21 11:07:06,602][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419411 seconds
+[2025-12-21 11:07:06,602][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442428 seconds
+[2025-12-21 11:07:07,073][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916226 seconds
+[2025-12-21 11:07:12,010][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999368 seconds
+[2025-12-21 11:07:12,025][openai._base_client][INFO] - Retrying request to /chat/completions in 0.896294 seconds
+[2025-12-21 11:07:12,025][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860609 seconds
+[2025-12-21 11:07:12,045][openai._base_client][INFO] - Retrying request to /chat/completions in 0.795295 seconds
+[2025-12-21 11:07:12,046][openai._base_client][INFO] - Retrying request to /chat/completions in 0.801173 seconds
+[2025-12-21 11:07:12,066][openai._base_client][INFO] - Retrying request to /chat/completions in 0.819079 seconds
+[2025-12-21 11:07:12,070][openai._base_client][INFO] - Retrying request to /chat/completions in 0.764586 seconds
+[2025-12-21 11:07:12,079][openai._base_client][INFO] - Retrying request to /chat/completions in 0.799439 seconds
+[2025-12-21 11:07:28,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:30,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:30,503][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:30,525][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:40,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:41,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:41,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:42,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:42,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:07:47,313][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439576 seconds
+[2025-12-21 11:07:52,328][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402919 seconds
+[2025-12-21 11:07:52,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413118 seconds
+[2025-12-21 11:07:52,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420189 seconds
+[2025-12-21 11:07:52,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391150 seconds
+[2025-12-21 11:07:52,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496677 seconds
+[2025-12-21 11:07:52,756][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767075 seconds
+[2025-12-21 11:07:53,101][openai._base_client][INFO] - Retrying request to /chat/completions in 0.910504 seconds
+[2025-12-21 11:07:53,128][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863663 seconds
+[2025-12-21 11:07:53,132][openai._base_client][INFO] - Retrying request to /chat/completions in 0.812341 seconds
+[2025-12-21 11:07:53,262][openai._base_client][INFO] - Retrying request to /chat/completions in 0.752987 seconds
+[2025-12-21 11:07:57,750][openai._base_client][INFO] - Retrying request to /chat/completions in 0.829584 seconds
+[2025-12-21 11:08:03,650][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:08:05,745][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:08:14,521][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:08:15,113][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:08:15,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:08:20,417][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380523 seconds
+[2025-12-21 11:08:20,801][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429956 seconds
+[2025-12-21 11:08:20,804][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431680 seconds
+[2025-12-21 11:08:20,816][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449108 seconds
+[2025-12-21 11:08:25,431][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407514 seconds
+[2025-12-21 11:08:25,432][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410844 seconds
+[2025-12-21 11:08:25,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432098 seconds
+[2025-12-21 11:08:25,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469051 seconds
+[2025-12-21 11:08:25,433][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472704 seconds
+[2025-12-21 11:08:25,804][openai._base_client][INFO] - Retrying request to /chat/completions in 0.912495 seconds
+[2025-12-21 11:08:26,235][openai._base_client][INFO] - Retrying request to /chat/completions in 0.872593 seconds
+[2025-12-21 11:08:26,238][openai._base_client][INFO] - Retrying request to /chat/completions in 0.787159 seconds
+[2025-12-21 11:08:26,269][openai._base_client][INFO] - Retrying request to /chat/completions in 0.946146 seconds
+[2025-12-21 11:08:30,846][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833384 seconds
+[2025-12-21 11:08:30,847][openai._base_client][INFO] - Retrying request to /chat/completions in 0.953354 seconds
+[2025-12-21 11:08:30,868][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916910 seconds
+[2025-12-21 11:08:30,906][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908303 seconds
+[2025-12-21 11:08:30,907][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842710 seconds
+[2025-12-21 11:08:46,838][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446025 seconds
+[2025-12-21 11:08:46,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441098 seconds
+[2025-12-21 11:08:46,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490675 seconds
+[2025-12-21 11:08:46,839][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430680 seconds
+[2025-12-21 11:08:46,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447288 seconds
+[2025-12-21 11:08:46,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405911 seconds
+[2025-12-21 11:08:46,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399235 seconds
+[2025-12-21 11:08:46,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450239 seconds
+[2025-12-21 11:08:46,840][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438424 seconds
+[2025-12-21 11:08:52,245][openai._base_client][INFO] - Retrying request to /chat/completions in 0.913741 seconds
+[2025-12-21 11:08:52,249][openai._base_client][INFO] - Retrying request to /chat/completions in 0.978227 seconds
+[2025-12-21 11:08:52,273][openai._base_client][INFO] - Retrying request to /chat/completions in 0.851855 seconds
+[2025-12-21 11:08:52,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.862675 seconds
+[2025-12-21 11:08:52,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844426 seconds
+[2025-12-21 11:08:52,287][openai._base_client][INFO] - Retrying request to /chat/completions in 0.855089 seconds
+[2025-12-21 11:08:52,289][openai._base_client][INFO] - Retrying request to /chat/completions in 0.872071 seconds
+[2025-12-21 11:08:52,292][openai._base_client][INFO] - Retrying request to /chat/completions in 0.901232 seconds
+[2025-12-21 11:08:52,332][openai._base_client][INFO] - Retrying request to /chat/completions in 0.823170 seconds
+[2025-12-21 11:09:20,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:21,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:22,146][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:22,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:23,553][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:28,562][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468176 seconds
+[2025-12-21 11:09:34,039][openai._base_client][INFO] - Retrying request to /chat/completions in 0.865791 seconds
+[2025-12-21 11:09:44,933][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:55,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:55,948][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:56,643][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:56,885][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:57,020][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:57,088][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:09:57,425][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/11-11-03/.hydra/config.yaml b/outputs/2025-12-21/11-11-03/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..e07d7a42d04f7550198c6b91aa59c68adfd404dd
--- /dev/null
+++ b/outputs/2025-12-21/11-11-03/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 64000
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/11-11-03/.hydra/hydra.yaml b/outputs/2025-12-21/11-11-03/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..911e96c5e0069e01180a76c67055c0a33c5b62ea
--- /dev/null
+++ b/outputs/2025-12-21/11-11-03/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/11-11-03
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/11-11-03/.hydra/overrides.yaml b/outputs/2025-12-21/11-11-03/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/11-11-03/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/11-11-03/eval_api.log b/outputs/2025-12-21/11-11-03/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..0284061ffce5df5721ff3f03396c9d53953aea50
--- /dev/null
+++ b/outputs/2025-12-21/11-11-03/eval_api.log
@@ -0,0 +1,346 @@
+[2025-12-21 11:11:10,202][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429283 seconds
+[2025-12-21 11:11:10,203][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458864 seconds
+[2025-12-21 11:11:10,203][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483376 seconds
+[2025-12-21 11:11:10,203][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450838 seconds
+[2025-12-21 11:11:10,204][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385313 seconds
+[2025-12-21 11:11:10,204][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414376 seconds
+[2025-12-21 11:11:10,205][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450111 seconds
+[2025-12-21 11:11:10,205][openai._base_client][INFO] - Retrying request to /chat/completions in 0.491654 seconds
+[2025-12-21 11:11:10,205][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430700 seconds
+[2025-12-21 11:11:10,205][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468063 seconds
+[2025-12-21 11:11:15,625][openai._base_client][INFO] - Retrying request to /chat/completions in 0.909452 seconds
+[2025-12-21 11:11:15,638][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833588 seconds
+[2025-12-21 11:11:15,660][openai._base_client][INFO] - Retrying request to /chat/completions in 0.923259 seconds
+[2025-12-21 11:11:15,666][openai._base_client][INFO] - Retrying request to /chat/completions in 0.908465 seconds
+[2025-12-21 11:11:15,690][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944448 seconds
+[2025-12-21 11:11:20,788][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:21,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:21,888][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:22,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:25,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:26,478][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455331 seconds
+[2025-12-21 11:11:26,542][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490460 seconds
+[2025-12-21 11:11:26,640][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494722 seconds
+[2025-12-21 11:11:27,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:31,083][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:32,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949425 seconds
+[2025-12-21 11:11:33,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:34,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:38,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:39,126][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:41,716][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:41,746][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:43,824][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:44,364][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:44,696][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:49,944][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:50,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:51,566][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:11:56,851][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:12:01,477][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:12:12,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:12:23,285][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:12:40,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:12:43,711][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:13:00,478][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:13:21,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:13:31,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:13:36,074][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422150 seconds
+[2025-12-21 11:13:41,081][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426839 seconds
+[2025-12-21 11:13:41,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430095 seconds
+[2025-12-21 11:13:46,516][openai._base_client][INFO] - Retrying request to /chat/completions in 0.910640 seconds
+[2025-12-21 11:13:46,517][openai._base_client][INFO] - Retrying request to /chat/completions in 0.935523 seconds
+[2025-12-21 11:13:46,838][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:04,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:11,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:16,162][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481136 seconds
+[2025-12-21 11:14:21,652][openai._base_client][INFO] - Retrying request to /chat/completions in 0.832270 seconds
+[2025-12-21 11:14:36,453][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:36,680][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375009 seconds
+[2025-12-21 11:14:41,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:41,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:41,649][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:41,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409919 seconds
+[2025-12-21 11:14:41,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.398447 seconds
+[2025-12-21 11:14:41,683][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423337 seconds
+[2025-12-21 11:14:41,713][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:41,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:42,141][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:42,922][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:43,064][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:43,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:43,744][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:43,887][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:44,965][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:45,434][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:46,656][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:47,094][openai._base_client][INFO] - Retrying request to /chat/completions in 0.799129 seconds
+[2025-12-21 11:14:47,109][openai._base_client][INFO] - Retrying request to /chat/completions in 0.931990 seconds
+[2025-12-21 11:14:47,253][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:47,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:47,727][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:47,825][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:48,677][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:49,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:49,948][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:50,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:51,053][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:51,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:51,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:52,357][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:52,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:52,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:52,837][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:53,010][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:55,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:57,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:14:57,253][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389488 seconds
+[2025-12-21 11:14:57,254][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394944 seconds
+[2025-12-21 11:14:57,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459380 seconds
+[2025-12-21 11:14:57,257][openai._base_client][INFO] - Retrying request to /chat/completions in 0.477861 seconds
+[2025-12-21 11:14:57,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460119 seconds
+[2025-12-21 11:14:57,260][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394021 seconds
+[2025-12-21 11:14:57,262][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431945 seconds
+[2025-12-21 11:14:59,846][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:00,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:01,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:01,554][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:01,775][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:02,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:02,262][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429186 seconds
+[2025-12-21 11:15:02,263][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383810 seconds
+[2025-12-21 11:15:02,263][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463339 seconds
+[2025-12-21 11:15:02,396][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:02,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:02,696][openai._base_client][INFO] - Retrying request to /chat/completions in 0.766436 seconds
+[2025-12-21 11:15:02,721][openai._base_client][INFO] - Retrying request to /chat/completions in 0.921395 seconds
+[2025-12-21 11:15:03,748][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:03,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:05,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:05,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:05,784][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:06,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:06,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:06,541][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:06,559][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:06,905][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:06,919][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:07,628][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:07,825][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:08,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:09,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:09,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:09,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:11,433][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:12,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:13,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:35,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:40,297][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466571 seconds
+[2025-12-21 11:15:51,805][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:56,215][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:56,456][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410677 seconds
+[2025-12-21 11:15:56,457][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396159 seconds
+[2025-12-21 11:15:59,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:59,513][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:15:59,583][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:00,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:00,378][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:01,141][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:01,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:01,458][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413391 seconds
+[2025-12-21 11:16:01,458][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396294 seconds
+[2025-12-21 11:16:01,460][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447932 seconds
+[2025-12-21 11:16:01,461][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486457 seconds
+[2025-12-21 11:16:01,461][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441958 seconds
+[2025-12-21 11:16:01,803][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:01,856][openai._base_client][INFO] - Retrying request to /chat/completions in 0.850718 seconds
+[2025-12-21 11:16:01,869][openai._base_client][INFO] - Retrying request to /chat/completions in 0.836449 seconds
+[2025-12-21 11:16:01,986][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:02,028][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:02,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:02,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:02,782][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:04,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:04,277][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:04,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:05,212][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:05,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:05,646][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:06,493][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:06,858][openai._base_client][INFO] - Retrying request to /chat/completions in 0.948525 seconds
+[2025-12-21 11:16:06,874][openai._base_client][INFO] - Retrying request to /chat/completions in 0.905961 seconds
+[2025-12-21 11:16:06,906][openai._base_client][INFO] - Retrying request to /chat/completions in 0.899903 seconds
+[2025-12-21 11:16:07,002][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:08,267][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:09,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:10,327][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:10,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:11,493][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:16,503][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486570 seconds
+[2025-12-21 11:16:21,509][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461000 seconds
+[2025-12-21 11:16:22,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:24,485][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:24,742][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424749 seconds
+[2025-12-21 11:16:24,743][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408670 seconds
+[2025-12-21 11:16:27,381][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:27,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:27,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:27,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:27,750][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:27,751][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:27,967][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:28,146][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:28,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:28,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:28,975][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:29,002][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:29,068][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:29,508][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:29,744][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419602 seconds
+[2025-12-21 11:16:29,745][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469305 seconds
+[2025-12-21 11:16:29,746][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437988 seconds
+[2025-12-21 11:16:30,885][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:30,984][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:31,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:31,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:31,820][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:32,676][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:32,689][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:32,840][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:32,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:33,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:46,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:46,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419058 seconds
+[2025-12-21 11:16:49,003][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:49,023][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:49,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:49,046][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:49,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:49,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:49,745][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:49,960][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:50,085][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:50,478][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:50,671][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:50,796][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:51,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:51,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:51,607][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443630 seconds
+[2025-12-21 11:16:52,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:53,693][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:53,763][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:53,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:55,252][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:57,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:16:57,324][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439528 seconds
+[2025-12-21 11:16:57,325][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473627 seconds
+[2025-12-21 11:16:57,326][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457727 seconds
+[2025-12-21 11:16:57,327][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434599 seconds
+[2025-12-21 11:16:57,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480272 seconds
+[2025-12-21 11:16:57,333][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460357 seconds
+[2025-12-21 11:16:59,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:00,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:01,301][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:01,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:02,180][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:02,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:02,307][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:02,329][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458887 seconds
+[2025-12-21 11:17:02,330][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457261 seconds
+[2025-12-21 11:17:02,505][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:02,752][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:02,765][openai._base_client][INFO] - Retrying request to /chat/completions in 0.857318 seconds
+[2025-12-21 11:17:02,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.859985 seconds
+[2025-12-21 11:17:02,786][openai._base_client][INFO] - Retrying request to /chat/completions in 0.920193 seconds
+[2025-12-21 11:17:02,812][openai._base_client][INFO] - Retrying request to /chat/completions in 0.900681 seconds
+[2025-12-21 11:17:03,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:05,459][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:05,482][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:06,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:06,049][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:06,185][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:06,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:07,763][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:07,959][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454851 seconds
+[2025-12-21 11:17:07,960][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409806 seconds
+[2025-12-21 11:17:07,962][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462438 seconds
+[2025-12-21 11:17:07,963][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498982 seconds
+[2025-12-21 11:17:07,964][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401644 seconds
+[2025-12-21 11:17:07,965][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427061 seconds
+[2025-12-21 11:17:07,979][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489628 seconds
+[2025-12-21 11:17:07,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481465 seconds
+[2025-12-21 11:17:10,558][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:11,117][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:11,217][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:11,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:12,968][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488017 seconds
+[2025-12-21 11:17:12,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422942 seconds
+[2025-12-21 11:17:12,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422091 seconds
+[2025-12-21 11:17:12,981][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417444 seconds
+[2025-12-21 11:17:13,370][openai._base_client][INFO] - Retrying request to /chat/completions in 0.883434 seconds
+[2025-12-21 11:17:13,373][openai._base_client][INFO] - Retrying request to /chat/completions in 0.766883 seconds
+[2025-12-21 11:17:13,395][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864694 seconds
+[2025-12-21 11:17:13,416][openai._base_client][INFO] - Retrying request to /chat/completions in 0.765206 seconds
+[2025-12-21 11:17:13,464][openai._base_client][INFO] - Retrying request to /chat/completions in 0.966954 seconds
+[2025-12-21 11:17:13,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.846649 seconds
+[2025-12-21 11:17:15,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:16,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:16,540][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:21,862][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:51,863][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:17:56,881][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479368 seconds
+[2025-12-21 11:17:59,649][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:00,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:01,897][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466217 seconds
+[2025-12-21 11:18:01,897][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402070 seconds
+[2025-12-21 11:18:01,898][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483797 seconds
+[2025-12-21 11:18:04,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:05,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:13,517][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:13,771][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427957 seconds
+[2025-12-21 11:18:16,290][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:16,416][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:16,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:16,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:16,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:16,952][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:17,272][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:17,833][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:18,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:19,033][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468791 seconds
+[2025-12-21 11:18:19,516][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:19,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:24,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:31,109][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:35,688][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:35,946][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382222 seconds
+[2025-12-21 11:18:35,947][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413063 seconds
+[2025-12-21 11:18:38,624][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:39,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:40,567][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:40,922][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:40,947][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395280 seconds
+[2025-12-21 11:18:40,948][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467974 seconds
+[2025-12-21 11:18:40,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422328 seconds
+[2025-12-21 11:18:40,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388197 seconds
+[2025-12-21 11:18:40,951][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488849 seconds
+[2025-12-21 11:18:40,951][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402110 seconds
+[2025-12-21 11:18:41,331][openai._base_client][INFO] - Retrying request to /chat/completions in 0.756735 seconds
+[2025-12-21 11:18:41,363][openai._base_client][INFO] - Retrying request to /chat/completions in 0.993086 seconds
+[2025-12-21 11:18:42,142][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:42,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:46,156][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:46,238][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:46,419][openai._base_client][INFO] - Retrying request to /chat/completions in 0.897642 seconds
+[2025-12-21 11:18:46,443][openai._base_client][INFO] - Retrying request to /chat/completions in 0.784066 seconds
+[2025-12-21 11:18:47,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:47,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:49,652][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:51,033][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:18:55,954][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:19:00,963][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488171 seconds
+[2025-12-21 11:19:06,591][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:19:06,657][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383706 seconds
+[2025-12-21 11:19:09,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:19:12,045][openai._base_client][INFO] - Retrying request to /chat/completions in 0.878960 seconds
+[2025-12-21 11:19:17,822][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/11-21-45/.hydra/config.yaml b/outputs/2025-12-21/11-21-45/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..a546d6db5f0f3752e4d8ac4a5339dfb0f3c88d06
--- /dev/null
+++ b/outputs/2025-12-21/11-21-45/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 6
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 64000
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/11-21-45/.hydra/hydra.yaml b/outputs/2025-12-21/11-21-45/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..aabb2514edd71c4bfb333e553d123c5f0d1a0efd
--- /dev/null
+++ b/outputs/2025-12-21/11-21-45/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/11-21-45
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/11-21-45/.hydra/overrides.yaml b/outputs/2025-12-21/11-21-45/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/11-21-45/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/11-21-45/eval_api.log b/outputs/2025-12-21/11-21-45/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..46cc2dacaff84273cc3b020de192888e8e63fdb7
--- /dev/null
+++ b/outputs/2025-12-21/11-21-45/eval_api.log
@@ -0,0 +1,233 @@
+[2025-12-21 11:21:52,255][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406878 seconds
+[2025-12-21 11:21:52,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405895 seconds
+[2025-12-21 11:21:57,669][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806388 seconds
+[2025-12-21 11:21:57,670][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844787 seconds
+[2025-12-21 11:22:02,349][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:22:09,892][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:22:27,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:22:41,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:22:43,679][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:22:46,855][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:22:47,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:22:49,058][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:01,818][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:04,265][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:14,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:19,594][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:21,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:24,590][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:32,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:32,672][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:33,379][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:34,295][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:40,243][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:40,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:41,697][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:47,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:49,690][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:50,218][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:52,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:23:53,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:24:02,954][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:24:22,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:24:24,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:24:50,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:24:56,935][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:04,327][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:04,454][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407447 seconds
+[2025-12-21 11:25:09,459][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485387 seconds
+[2025-12-21 11:25:09,865][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781075 seconds
+[2025-12-21 11:25:13,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:14,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.898207 seconds
+[2025-12-21 11:25:16,152][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:20,658][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395792 seconds
+[2025-12-21 11:25:21,968][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:22,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:23,793][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:25,229][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:26,059][openai._base_client][INFO] - Retrying request to /chat/completions in 0.853664 seconds
+[2025-12-21 11:25:28,026][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:29,274][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:29,512][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:33,199][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:34,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:37,272][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:42,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:43,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:45,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:51,197][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:52,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:53,285][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:53,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:55,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:25:55,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:26:13,521][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:26:20,006][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:26:25,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388594 seconds
+[2025-12-21 11:26:30,049][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409644 seconds
+[2025-12-21 11:26:35,467][openai._base_client][INFO] - Retrying request to /chat/completions in 0.754092 seconds
+[2025-12-21 11:27:09,925][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:27:27,270][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:27:32,278][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376267 seconds
+[2025-12-21 11:27:55,458][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:27:55,580][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480230 seconds
+[2025-12-21 11:27:59,784][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:00,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419916 seconds
+[2025-12-21 11:28:00,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497345 seconds
+[2025-12-21 11:28:00,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478463 seconds
+[2025-12-21 11:28:00,583][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496348 seconds
+[2025-12-21 11:28:00,932][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:01,063][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869588 seconds
+[2025-12-21 11:28:06,007][openai._base_client][INFO] - Retrying request to /chat/completions in 0.872058 seconds
+[2025-12-21 11:28:06,064][openai._base_client][INFO] - Retrying request to /chat/completions in 0.772465 seconds
+[2025-12-21 11:28:06,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.914306 seconds
+[2025-12-21 11:28:06,082][openai._base_client][INFO] - Retrying request to /chat/completions in 0.915089 seconds
+[2025-12-21 11:28:06,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:10,715][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:11,941][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444288 seconds
+[2025-12-21 11:28:12,669][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:16,844][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452468 seconds
+[2025-12-21 11:28:16,886][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463457 seconds
+[2025-12-21 11:28:16,887][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:17,006][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388811 seconds
+[2025-12-21 11:28:17,388][openai._base_client][INFO] - Retrying request to /chat/completions in 0.956099 seconds
+[2025-12-21 11:28:18,843][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:22,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:22,353][openai._base_client][INFO] - Retrying request to /chat/completions in 0.752539 seconds
+[2025-12-21 11:28:22,398][openai._base_client][INFO] - Retrying request to /chat/completions in 0.947619 seconds
+[2025-12-21 11:28:24,298][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:24,325][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:26,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:26,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:27,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:27,495][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:27,561][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:33,227][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:33,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:33,765][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:40,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:28:55,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:29:00,883][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422799 seconds
+[2025-12-21 11:29:05,900][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450373 seconds
+[2025-12-21 11:29:05,901][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375748 seconds
+[2025-12-21 11:29:05,902][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476901 seconds
+[2025-12-21 11:29:05,902][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396590 seconds
+[2025-12-21 11:29:06,313][openai._base_client][INFO] - Retrying request to /chat/completions in 0.987964 seconds
+[2025-12-21 11:29:11,361][openai._base_client][INFO] - Retrying request to /chat/completions in 0.940059 seconds
+[2025-12-21 11:29:17,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:29:27,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:29:47,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:29:51,685][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:29:57,993][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:03,002][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478465 seconds
+[2025-12-21 11:30:23,373][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:23,520][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462926 seconds
+[2025-12-21 11:30:28,433][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:28,520][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453986 seconds
+[2025-12-21 11:30:28,521][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409219 seconds
+[2025-12-21 11:30:28,521][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386647 seconds
+[2025-12-21 11:30:28,522][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454963 seconds
+[2025-12-21 11:30:28,522][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402433 seconds
+[2025-12-21 11:30:28,522][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394346 seconds
+[2025-12-21 11:30:28,523][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416350 seconds
+[2025-12-21 11:30:29,094][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:29,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:30,545][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:32,406][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:33,069][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:33,682][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:33,941][openai._base_client][INFO] - Retrying request to /chat/completions in 0.760837 seconds
+[2025-12-21 11:30:34,898][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:35,189][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:36,046][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:36,461][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:37,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:46,276][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:46,648][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:48,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:49,172][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:49,575][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:50,217][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:30:55,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:05,296][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:07,656][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:14,372][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:19,384][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459034 seconds
+[2025-12-21 11:31:25,707][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:25,911][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450270 seconds
+[2025-12-21 11:31:29,183][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:29,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:30,914][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492838 seconds
+[2025-12-21 11:31:30,919][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452055 seconds
+[2025-12-21 11:31:30,920][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492482 seconds
+[2025-12-21 11:31:30,926][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375653 seconds
+[2025-12-21 11:31:30,930][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447359 seconds
+[2025-12-21 11:31:30,935][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446767 seconds
+[2025-12-21 11:31:30,935][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499226 seconds
+[2025-12-21 11:31:30,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496172 seconds
+[2025-12-21 11:31:31,365][openai._base_client][INFO] - Retrying request to /chat/completions in 0.766082 seconds
+[2025-12-21 11:31:31,391][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:31,696][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:32,796][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:33,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:34,946][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:35,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:36,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:36,307][openai._base_client][INFO] - Retrying request to /chat/completions in 0.820189 seconds
+[2025-12-21 11:31:36,374][openai._base_client][INFO] - Retrying request to /chat/completions in 0.949953 seconds
+[2025-12-21 11:31:36,385][openai._base_client][INFO] - Retrying request to /chat/completions in 0.794708 seconds
+[2025-12-21 11:31:36,385][openai._base_client][INFO] - Retrying request to /chat/completions in 0.950324 seconds
+[2025-12-21 11:31:36,413][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863564 seconds
+[2025-12-21 11:31:36,420][openai._base_client][INFO] - Retrying request to /chat/completions in 0.922377 seconds
+[2025-12-21 11:31:36,437][openai._base_client][INFO] - Retrying request to /chat/completions in 0.816255 seconds
+[2025-12-21 11:31:36,438][openai._base_client][INFO] - Retrying request to /chat/completions in 0.924988 seconds
+[2025-12-21 11:31:36,768][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:37,970][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:42,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:31:46,502][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:32:10,675][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:32:21,113][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:32:28,141][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:32:29,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:32:34,866][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379837 seconds
+[2025-12-21 11:32:39,883][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463576 seconds
+[2025-12-21 11:32:39,890][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473815 seconds
+[2025-12-21 11:32:39,891][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425410 seconds
+[2025-12-21 11:32:39,892][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492239 seconds
+[2025-12-21 11:32:39,892][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460560 seconds
+[2025-12-21 11:32:40,250][openai._base_client][INFO] - Retrying request to /chat/completions in 0.834970 seconds
+[2025-12-21 11:32:44,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:32:45,359][openai._base_client][INFO] - Retrying request to /chat/completions in 0.945493 seconds
+[2025-12-21 11:32:45,387][openai._base_client][INFO] - Retrying request to /chat/completions in 0.829527 seconds
+[2025-12-21 11:32:46,795][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:32:56,520][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:33:00,158][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:34:01,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:34:06,547][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437589 seconds
+[2025-12-21 11:34:11,992][openai._base_client][INFO] - Retrying request to /chat/completions in 0.846617 seconds
+[2025-12-21 11:34:53,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:34:54,234][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461452 seconds
+[2025-12-21 11:34:54,535][openai._base_client][INFO] - Retrying request to /chat/completions in 0.415802 seconds
+[2025-12-21 11:34:55,999][openai._base_client][INFO] - Retrying request to /chat/completions in 0.845765 seconds
+[2025-12-21 11:34:58,499][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:34:58,777][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:34:59,025][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:34:59,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:34:59,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:00,173][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:00,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:00,763][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:01,595][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:01,910][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:02,084][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:02,232][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:03,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:03,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:03,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:03,941][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:04,131][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:04,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:06,011][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:10,075][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:22,736][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:35:23,778][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/11-37-15/.hydra/config.yaml b/outputs/2025-12-21/11-37-15/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..514c365f6e0da8fd353cfd6d51f9ccb6bdf47474
--- /dev/null
+++ b/outputs/2025-12-21/11-37-15/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 2
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 6
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 64000
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/11-37-15/.hydra/hydra.yaml b/outputs/2025-12-21/11-37-15/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..17cdd4387a1dc82ab292681cb40662137430f9b6
--- /dev/null
+++ b/outputs/2025-12-21/11-37-15/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/11-37-15
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/11-37-15/.hydra/overrides.yaml b/outputs/2025-12-21/11-37-15/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/11-37-15/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/11-37-15/eval_api.log b/outputs/2025-12-21/11-37-15/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..4aa671263760f26bd9f9bb93dcc8d39f3fcbe4b2
--- /dev/null
+++ b/outputs/2025-12-21/11-37-15/eval_api.log
@@ -0,0 +1,298 @@
+[2025-12-21 11:37:16,994][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401147 seconds
+[2025-12-21 11:37:16,994][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376829 seconds
+[2025-12-21 11:37:16,995][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420802 seconds
+[2025-12-21 11:37:17,043][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481892 seconds
+[2025-12-21 11:37:17,044][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464993 seconds
+[2025-12-21 11:37:17,047][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401242 seconds
+[2025-12-21 11:37:17,381][openai._base_client][INFO] - Retrying request to /chat/completions in 0.880293 seconds
+[2025-12-21 11:37:17,519][openai._base_client][INFO] - Retrying request to /chat/completions in 0.882871 seconds
+[2025-12-21 11:37:17,534][openai._base_client][INFO] - Retrying request to /chat/completions in 0.809254 seconds
+[2025-12-21 11:37:18,280][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474806 seconds
+[2025-12-21 11:37:18,361][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421649 seconds
+[2025-12-21 11:37:18,423][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489481 seconds
+[2025-12-21 11:37:18,764][openai._base_client][INFO] - Retrying request to /chat/completions in 0.928762 seconds
+[2025-12-21 11:37:18,793][openai._base_client][INFO] - Retrying request to /chat/completions in 0.983859 seconds
+[2025-12-21 11:37:18,924][openai._base_client][INFO] - Retrying request to /chat/completions in 0.976779 seconds
+[2025-12-21 11:37:19,711][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435508 seconds
+[2025-12-21 11:37:19,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382583 seconds
+[2025-12-21 11:37:19,921][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375723 seconds
+[2025-12-21 11:37:20,188][openai._base_client][INFO] - Retrying request to /chat/completions in 0.781648 seconds
+[2025-12-21 11:37:20,307][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818793 seconds
+[2025-12-21 11:37:29,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:37:30,417][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:37:34,292][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:37:35,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:37:44,052][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:37:47,645][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:37:54,410][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:37:55,448][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:01,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:02,051][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:14,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:16,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:31,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:32,895][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:35,789][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:40,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:47,107][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:54,070][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:38:58,408][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:03,474][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:03,488][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:03,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:07,680][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:11,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:22,439][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:44,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:39:49,667][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443592 seconds
+[2025-12-21 11:39:49,758][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475015 seconds
+[2025-12-21 11:39:54,680][openai._base_client][INFO] - Retrying request to /chat/completions in 0.399722 seconds
+[2025-12-21 11:39:55,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.893066 seconds
+[2025-12-21 11:39:55,237][openai._base_client][INFO] - Retrying request to /chat/completions in 0.808796 seconds
+[2025-12-21 11:40:00,085][openai._base_client][INFO] - Retrying request to /chat/completions in 0.756258 seconds
+[2025-12-21 11:40:02,515][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:40:27,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:41:29,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:41:34,829][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379424 seconds
+[2025-12-21 11:41:39,834][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496522 seconds
+[2025-12-21 11:41:39,836][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393143 seconds
+[2025-12-21 11:42:12,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:42:28,370][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:24,435][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:24,554][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386329 seconds
+[2025-12-21 11:43:29,543][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:29,558][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418469 seconds
+[2025-12-21 11:43:29,558][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463445 seconds
+[2025-12-21 11:43:29,558][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439250 seconds
+[2025-12-21 11:43:29,559][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467771 seconds
+[2025-12-21 11:43:29,893][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452524 seconds
+[2025-12-21 11:43:30,825][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:34,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:35,001][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933809 seconds
+[2025-12-21 11:43:35,024][openai._base_client][INFO] - Retrying request to /chat/completions in 0.818548 seconds
+[2025-12-21 11:43:35,349][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871186 seconds
+[2025-12-21 11:43:36,773][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:37,004][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:41,314][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:41,681][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:44,815][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:46,227][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455710 seconds
+[2025-12-21 11:43:46,280][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:49,766][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:51,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:51,687][openai._base_client][INFO] - Retrying request to /chat/completions in 0.977228 seconds
+[2025-12-21 11:43:52,262][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:52,306][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:53,655][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:55,461][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:43:57,291][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:01,207][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:04,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:05,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:05,541][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:11,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:15,998][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:17,418][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:18,759][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:20,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:27,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:35,482][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:44:58,215][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:53:12,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:53:17,638][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393368 seconds
+[2025-12-21 11:53:22,645][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481471 seconds
+[2025-12-21 11:53:22,647][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495646 seconds
+[2025-12-21 11:53:23,034][openai._base_client][INFO] - Retrying request to /chat/completions in 0.906759 seconds
+[2025-12-21 11:53:28,134][openai._base_client][INFO] - Retrying request to /chat/completions in 0.931341 seconds
+[2025-12-21 11:53:28,145][openai._base_client][INFO] - Retrying request to /chat/completions in 0.940837 seconds
+[2025-12-21 11:53:29,239][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:53:32,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:53:39,073][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454353 seconds
+[2025-12-21 11:55:12,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:12,745][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489722 seconds
+[2025-12-21 11:55:13,963][openai._base_client][INFO] - Retrying request to /chat/completions in 0.828755 seconds
+[2025-12-21 11:55:16,258][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:16,276][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378463 seconds
+[2025-12-21 11:55:17,747][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389182 seconds
+[2025-12-21 11:55:17,747][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470487 seconds
+[2025-12-21 11:55:17,747][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402972 seconds
+[2025-12-21 11:55:17,748][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412367 seconds
+[2025-12-21 11:55:17,748][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452654 seconds
+[2025-12-21 11:55:17,749][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396088 seconds
+[2025-12-21 11:55:17,749][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408322 seconds
+[2025-12-21 11:55:17,749][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434952 seconds
+[2025-12-21 11:55:17,749][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392693 seconds
+[2025-12-21 11:55:17,750][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418656 seconds
+[2025-12-21 11:55:17,751][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499666 seconds
+[2025-12-21 11:55:18,879][openai._base_client][INFO] - Retrying request to /chat/completions in 0.796669 seconds
+[2025-12-21 11:55:18,880][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842380 seconds
+[2025-12-21 11:55:18,888][openai._base_client][INFO] - Retrying request to /chat/completions in 0.769639 seconds
+[2025-12-21 11:55:21,659][openai._base_client][INFO] - Retrying request to /chat/completions in 0.843168 seconds
+[2025-12-21 11:55:23,157][openai._base_client][INFO] - Retrying request to /chat/completions in 0.924889 seconds
+[2025-12-21 11:55:23,171][openai._base_client][INFO] - Retrying request to /chat/completions in 0.755534 seconds
+[2025-12-21 11:55:23,187][openai._base_client][INFO] - Retrying request to /chat/completions in 0.996177 seconds
+[2025-12-21 11:55:23,204][openai._base_client][INFO] - Retrying request to /chat/completions in 0.943534 seconds
+[2025-12-21 11:55:23,221][openai._base_client][INFO] - Retrying request to /chat/completions in 0.764271 seconds
+[2025-12-21 11:55:23,734][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:24,396][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:26,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:27,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:29,190][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:34,047][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:37,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:38,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:43,360][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:44,962][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:45,929][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:51,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:55:51,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:00,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:02,058][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:02,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:18,872][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:20,773][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:29,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:34,817][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472711 seconds
+[2025-12-21 11:56:39,830][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375206 seconds
+[2025-12-21 11:56:49,446][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:57,582][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:56:57,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:22,007][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:29,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:31,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:31,588][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433950 seconds
+[2025-12-21 11:57:31,590][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473138 seconds
+[2025-12-21 11:57:36,592][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400922 seconds
+[2025-12-21 11:57:36,593][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383597 seconds
+[2025-12-21 11:57:36,593][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447531 seconds
+[2025-12-21 11:57:36,593][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381066 seconds
+[2025-12-21 11:57:36,593][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439479 seconds
+[2025-12-21 11:57:37,025][openai._base_client][INFO] - Retrying request to /chat/completions in 0.796467 seconds
+[2025-12-21 11:57:37,067][openai._base_client][INFO] - Retrying request to /chat/completions in 0.810680 seconds
+[2025-12-21 11:57:38,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:40,629][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:40,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:41,978][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815971 seconds
+[2025-12-21 11:57:41,979][openai._base_client][INFO] - Retrying request to /chat/completions in 0.992538 seconds
+[2025-12-21 11:57:41,995][openai._base_client][INFO] - Retrying request to /chat/completions in 0.958167 seconds
+[2025-12-21 11:57:42,035][openai._base_client][INFO] - Retrying request to /chat/completions in 0.972317 seconds
+[2025-12-21 11:57:42,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.939656 seconds
+[2025-12-21 11:57:44,481][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:44,604][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:45,563][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:46,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:47,121][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:47,828][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433255 seconds
+[2025-12-21 11:57:47,883][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391540 seconds
+[2025-12-21 11:57:48,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:49,619][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:53,281][openai._base_client][INFO] - Retrying request to /chat/completions in 0.779903 seconds
+[2025-12-21 11:57:53,487][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:55,680][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:57,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:57:57,983][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:58:01,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:58:05,387][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:58:37,166][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:59:04,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:59:09,689][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425551 seconds
+[2025-12-21 11:59:14,705][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449830 seconds
+[2025-12-21 11:59:14,706][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411294 seconds
+[2025-12-21 11:59:14,707][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401958 seconds
+[2025-12-21 11:59:14,707][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443160 seconds
+[2025-12-21 11:59:14,707][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453165 seconds
+[2025-12-21 11:59:14,708][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379486 seconds
+[2025-12-21 11:59:18,972][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:59:18,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:59:19,877][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:59:20,159][openai._base_client][INFO] - Retrying request to /chat/completions in 0.779241 seconds
+[2025-12-21 11:59:20,354][openai._base_client][INFO] - Retrying request to /chat/completions in 0.842742 seconds
+[2025-12-21 11:59:20,742][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:59:24,267][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 11:59:52,814][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:02,291][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:02,525][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429717 seconds
+[2025-12-21 12:00:06,626][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:06,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:06,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:07,150][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:07,528][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470404 seconds
+[2025-12-21 12:00:07,528][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379179 seconds
+[2025-12-21 12:00:07,529][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493139 seconds
+[2025-12-21 12:00:07,529][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452418 seconds
+[2025-12-21 12:00:08,231][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:08,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:09,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:09,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:10,373][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:12,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:12,444][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:13,004][openai._base_client][INFO] - Retrying request to /chat/completions in 0.975074 seconds
+[2025-12-21 12:00:13,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:14,035][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:15,159][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:15,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:16,333][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:16,364][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:18,174][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:24,118][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:26,022][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:31,408][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:35,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:47,125][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:58,207][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:00:58,271][root][WARNING] - Truncated history for env 2 to keep prompt within 7200 tokens (current 3353).
+[2025-12-21 12:00:58,447][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448782 seconds
+[2025-12-21 12:01:02,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:02,890][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:02,911][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:03,448][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449320 seconds
+[2025-12-21 12:01:03,449][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466754 seconds
+[2025-12-21 12:01:03,449][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437401 seconds
+[2025-12-21 12:01:03,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483581 seconds
+[2025-12-21 12:01:03,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457906 seconds
+[2025-12-21 12:01:03,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452703 seconds
+[2025-12-21 12:01:03,450][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495262 seconds
+[2025-12-21 12:01:03,451][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484783 seconds
+[2025-12-21 12:01:03,451][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492867 seconds
+[2025-12-21 12:01:03,452][openai._base_client][INFO] - Retrying request to /chat/completions in 0.392399 seconds
+[2025-12-21 12:01:03,899][openai._base_client][INFO] - Retrying request to /chat/completions in 0.760355 seconds
+[2025-12-21 12:01:04,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:04,224][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:04,952][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:05,468][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:06,611][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:07,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:08,579][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:08,849][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788155 seconds
+[2025-12-21 12:01:08,890][openai._base_client][INFO] - Retrying request to /chat/completions in 0.963208 seconds
+[2025-12-21 12:01:08,900][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797294 seconds
+[2025-12-21 12:01:08,905][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871915 seconds
+[2025-12-21 12:01:08,910][openai._base_client][INFO] - Retrying request to /chat/completions in 0.756946 seconds
+[2025-12-21 12:01:08,917][openai._base_client][INFO] - Retrying request to /chat/completions in 0.875211 seconds
+[2025-12-21 12:01:08,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.882631 seconds
+[2025-12-21 12:01:08,937][openai._base_client][INFO] - Retrying request to /chat/completions in 0.805379 seconds
+[2025-12-21 12:01:08,946][openai._base_client][INFO] - Retrying request to /chat/completions in 0.990911 seconds
+[2025-12-21 12:01:08,949][openai._base_client][INFO] - Retrying request to /chat/completions in 0.930688 seconds
+[2025-12-21 12:01:09,088][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:09,794][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:10,720][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:12,726][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:12,745][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:12,995][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:22,614][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:27,626][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456760 seconds
+[2025-12-21 12:01:31,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:32,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:32,643][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465532 seconds
+[2025-12-21 12:01:32,644][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480457 seconds
+[2025-12-21 12:01:32,644][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395049 seconds
+[2025-12-21 12:01:32,644][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437922 seconds
+[2025-12-21 12:01:35,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:36,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:38,114][openai._base_client][INFO] - Retrying request to /chat/completions in 0.892901 seconds
+[2025-12-21 12:01:38,127][openai._base_client][INFO] - Retrying request to /chat/completions in 0.826767 seconds
+[2025-12-21 12:01:41,184][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:01:57,245][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:02:13,036][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:02:13,194][root][WARNING] - Truncated history for env 14 to keep prompt within 7200 tokens (current 6968).
diff --git a/outputs/2025-12-21/11-50-29/.hydra/config.yaml b/outputs/2025-12-21/11-50-29/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8b212ba18f8827e7f22cbbf82cdc4db6ba7109f7
--- /dev/null
+++ b/outputs/2025-12-21/11-50-29/.hydra/config.yaml
@@ -0,0 +1,859 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 500
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_25_3b_it
+  experiment_name: rubikscube-3-starpos-SPA
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState
+      representation format: A string of colors representing faces U, D, L, R, F,
+      B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+      String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+      B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating
+      U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>'
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-3B-text_with_coordinates-qwen/global_step_1335
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 4
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-21/11-50-29/.hydra/hydra.yaml b/outputs/2025-12-21/11-50-29/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4cda49467bf54f8fe50b1faa1f2703ecd987a171
--- /dev/null
+++ b/outputs/2025-12-21/11-50-29/.hydra/hydra.yaml
@@ -0,0 +1,217 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_25_3b_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-3B-text_with_coordinates-qwen/global_step_1335
+    - custom_envs.rubikscube.env_config.scramble_depth=3
+    - 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube
+      (Pocket Cube).\nThe goal is to restore the cube so that each of the faces consists
+      of a single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'',
+      F, F'', B, B''.\nState representation format: A string of colors representing
+      faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+    - trainer.experiment_name=rubikscube-3-starpos-SPA
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.rubikscube.env_config.scramble_depth=3,custom_envs.rubikscube.env_instruction="You
+      are solving a 2x2 Rubik''s Cube (Pocket Cube).\nThe goal is to restore the cube
+      so that each of the faces consists of a single, unique color.\nAvailable actions:
+      U, U'', D, D'', L, L'', R, R'', F, F'', B, B''.\nState representation format:
+      A string of colors representing faces U, D, L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent
+      State: [Color String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace
+      R (Right): B B O O\n...\n</observation>\nI need to fix the Front face top row.
+      Rotating U (Up) clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+      State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+      Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n      W W\n      W
+      W\nO O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace
+      U: All White. Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face
+      D: Yellow.\nCube is currently solved.\n</observation>\nTo disrupt the state
+      for demonstration, I will apply R (Right face clockwise).\nThis will move columns:
+      F right col -> U right col -> B left col (inverted) -> D right col -> F right
+      col.\n<prediction>\n      W Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W
+      W   B B\n      Y R\n      Y R\nFace R rotates clockwise (colors cycle). Right
+      column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>",model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-3B-text_with_coordinates-qwen/global_step_1335,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=rubikscube-3-starpos-SPA,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_25_3b_it'
+    id: ???
+    num: ???
+    config_name: _10_rubikscube
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/11-50-29
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/11-50-29/.hydra/overrides.yaml b/outputs/2025-12-21/11-50-29/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..87c005766613c41a1466acd90334181e00169c5f
--- /dev/null
+++ b/outputs/2025-12-21/11-50-29/.hydra/overrides.yaml
@@ -0,0 +1,29 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_25_3b_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/SPA/sftckpt/checkpoints_10_rubikscube-3B-text_with_coordinates-qwen/global_step_1335
+- custom_envs.rubikscube.env_config.scramble_depth=3
+- 'custom_envs.rubikscube.env_instruction="You are solving a 2x2 Rubik''s Cube (Pocket
+  Cube).\nThe goal is to restore the cube so that each of the faces consists of a
+  single, unique color.\nAvailable actions: U, U'', D, D'', L, L'', R, R'', F, F'',
+  B, B''.\nState representation format: A string of colors representing faces U, D,
+  L, R, F, B.\n\nExample answer format:\n<think>\n<observation>\nCurrent State: [Color
+  String or Grid]\nFace U (Up): W W W W\nFace F (Front): R R G G\nFace R (Right):
+  B B O O\n...\n</observation>\nI need to fix the Front face top row. Rotating U (Up)
+  clockwise will shift the top row of F, R, B, L to the left.\n<prediction>\nPredicted
+  State: [Predicted Color String or Grid]\nFace F top row becomes the colors from
+  Face R.\nFace R top row becomes the colors from Face B.\n...\n</prediction>\n</think>\n<answer>U</answer>\n\nA
+  sample full output is as follows:\n<think>\n<observation>\n      W W\n      W W\nO
+  O   G G   R R   B B\nO O   G G   R R   B B\n      Y Y\n      Y Y\nFace U: All White.
+  Face F: Green. Face R: Red. Face B: Blue. Face L: Orange. Face D: Yellow.\nCube
+  is currently solved.\n</observation>\nTo disrupt the state for demonstration, I
+  will apply R (Right face clockwise).\nThis will move columns: F right col -> U right
+  col -> B left col (inverted) -> D right col -> F right col.\n<prediction>\n      W
+  Y\n      W Y\nO O   G G   W W   B B\nO O   G G   W W   B B\n      Y R\n      Y R\nFace
+  R rotates clockwise (colors cycle). Right column of F becomes Yellow (from D).\n</prediction>\n</think>\n<answer>R</answer>"'
+- trainer.experiment_name=rubikscube-3-starpos-SPA
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-21/11-50-29/train.log b/outputs/2025-12-21/11-50-29/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-21/12-07-03/.hydra/config.yaml b/outputs/2025-12-21/12-07-03/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..22071f9cd99d571288690c9f22df5e7d071cd93f
--- /dev/null
+++ b/outputs/2025-12-21/12-07-03/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 6
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - rubikscube
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 64000
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/12-07-03/.hydra/hydra.yaml b/outputs/2025-12-21/12-07-03/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..278f906805281fabb543c61ad9cf850b2c4de9a5
--- /dev/null
+++ b/outputs/2025-12-21/12-07-03/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/12-07-03
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/12-07-03/.hydra/overrides.yaml b/outputs/2025-12-21/12-07-03/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/12-07-03/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/12-07-03/eval_api.log b/outputs/2025-12-21/12-07-03/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..08dd4a27187acadfe73b5db3858f7df40f439ccd
--- /dev/null
+++ b/outputs/2025-12-21/12-07-03/eval_api.log
@@ -0,0 +1,285 @@
+[2025-12-21 12:07:10,298][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387399 seconds
+[2025-12-21 12:07:10,299][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442340 seconds
+[2025-12-21 12:07:10,299][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486717 seconds
+[2025-12-21 12:07:10,300][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431297 seconds
+[2025-12-21 12:07:10,300][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391676 seconds
+[2025-12-21 12:07:10,300][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410737 seconds
+[2025-12-21 12:07:10,300][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456029 seconds
+[2025-12-21 12:07:10,302][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486326 seconds
+[2025-12-21 12:07:15,694][openai._base_client][INFO] - Retrying request to /chat/completions in 0.979112 seconds
+[2025-12-21 12:07:15,695][openai._base_client][INFO] - Retrying request to /chat/completions in 0.954252 seconds
+[2025-12-21 12:07:15,712][openai._base_client][INFO] - Retrying request to /chat/completions in 0.975066 seconds
+[2025-12-21 12:07:15,733][openai._base_client][INFO] - Retrying request to /chat/completions in 0.983244 seconds
+[2025-12-21 12:07:15,743][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783595 seconds
+[2025-12-21 12:07:15,759][openai._base_client][INFO] - Retrying request to /chat/completions in 0.758687 seconds
+[2025-12-21 12:07:15,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.881895 seconds
+[2025-12-21 12:07:15,791][openai._base_client][INFO] - Retrying request to /chat/completions in 0.933947 seconds
+[2025-12-21 12:07:26,532][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456849 seconds
+[2025-12-21 12:07:26,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433648 seconds
+[2025-12-21 12:07:26,656][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437693 seconds
+[2025-12-21 12:07:26,676][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447459 seconds
+[2025-12-21 12:07:26,676][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496028 seconds
+[2025-12-21 12:07:26,691][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485765 seconds
+[2025-12-21 12:07:26,722][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481153 seconds
+[2025-12-21 12:07:26,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429804 seconds
+[2025-12-21 12:07:30,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:07:31,972][openai._base_client][INFO] - Retrying request to /chat/completions in 0.785800 seconds
+[2025-12-21 12:07:31,993][openai._base_client][INFO] - Retrying request to /chat/completions in 0.874226 seconds
+[2025-12-21 12:07:32,097][openai._base_client][INFO] - Retrying request to /chat/completions in 0.995003 seconds
+[2025-12-21 12:07:32,125][openai._base_client][INFO] - Retrying request to /chat/completions in 0.862726 seconds
+[2025-12-21 12:07:32,161][openai._base_client][INFO] - Retrying request to /chat/completions in 0.860034 seconds
+[2025-12-21 12:07:32,175][openai._base_client][INFO] - Retrying request to /chat/completions in 0.870566 seconds
+[2025-12-21 12:07:32,179][openai._base_client][INFO] - Retrying request to /chat/completions in 0.894944 seconds
+[2025-12-21 12:07:32,206][openai._base_client][INFO] - Retrying request to /chat/completions in 0.958924 seconds
+[2025-12-21 12:07:36,593][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:07:42,767][openai._base_client][INFO] - Retrying request to /chat/completions in 0.394997 seconds
+[2025-12-21 12:07:42,888][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488969 seconds
+[2025-12-21 12:07:43,050][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438371 seconds
+[2025-12-21 12:07:43,079][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393520 seconds
+[2025-12-21 12:07:43,170][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481123 seconds
+[2025-12-21 12:07:43,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:07:48,170][openai._base_client][INFO] - Retrying request to /chat/completions in 0.946275 seconds
+[2025-12-21 12:07:48,476][openai._base_client][INFO] - Retrying request to /chat/completions in 0.906733 seconds
+[2025-12-21 12:07:48,491][openai._base_client][INFO] - Retrying request to /chat/completions in 0.795985 seconds
+[2025-12-21 12:07:57,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:05,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:09,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:35,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:41,242][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:41,368][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:43,073][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:43,789][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:46,263][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:47,281][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:08:56,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:09:07,074][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:09:11,992][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:09:22,436][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:09:27,446][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396918 seconds
+[2025-12-21 12:09:32,463][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406813 seconds
+[2025-12-21 12:09:32,464][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389591 seconds
+[2025-12-21 12:09:32,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436449 seconds
+[2025-12-21 12:09:32,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407482 seconds
+[2025-12-21 12:09:32,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447103 seconds
+[2025-12-21 12:09:32,465][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430264 seconds
+[2025-12-21 12:09:32,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447821 seconds
+[2025-12-21 12:09:32,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.458201 seconds
+[2025-12-21 12:09:32,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465683 seconds
+[2025-12-21 12:09:32,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431157 seconds
+[2025-12-21 12:09:32,466][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478687 seconds
+[2025-12-21 12:09:32,847][openai._base_client][INFO] - Retrying request to /chat/completions in 0.808306 seconds
+[2025-12-21 12:09:37,860][openai._base_client][INFO] - Retrying request to /chat/completions in 0.915869 seconds
+[2025-12-21 12:09:37,899][openai._base_client][INFO] - Retrying request to /chat/completions in 0.951360 seconds
+[2025-12-21 12:09:37,903][openai._base_client][INFO] - Retrying request to /chat/completions in 0.771467 seconds
+[2025-12-21 12:09:37,916][openai._base_client][INFO] - Retrying request to /chat/completions in 0.981876 seconds
+[2025-12-21 12:09:37,926][openai._base_client][INFO] - Retrying request to /chat/completions in 0.823966 seconds
+[2025-12-21 12:09:37,934][openai._base_client][INFO] - Retrying request to /chat/completions in 0.774953 seconds
+[2025-12-21 12:09:42,123][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:09:51,293][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:10:30,936][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:10:43,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:10:48,289][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:10:50,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:11:06,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:11:18,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:11:32,616][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:11:38,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:11:49,240][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:11:54,250][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486277 seconds
+[2025-12-21 12:11:59,259][openai._base_client][INFO] - Retrying request to /chat/completions in 0.485594 seconds
+[2025-12-21 12:11:59,260][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375128 seconds
+[2025-12-21 12:11:59,740][openai._base_client][INFO] - Retrying request to /chat/completions in 0.954598 seconds
+[2025-12-21 12:12:04,642][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844858 seconds
+[2025-12-21 12:12:04,749][openai._base_client][INFO] - Retrying request to /chat/completions in 0.751021 seconds
+[2025-12-21 12:12:35,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:13:11,235][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:13:16,244][openai._base_client][INFO] - Retrying request to /chat/completions in 0.411727 seconds
+[2025-12-21 12:13:21,251][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409586 seconds
+[2025-12-21 12:13:21,659][openai._base_client][INFO] - Retrying request to /chat/completions in 0.852414 seconds
+[2025-12-21 12:13:26,668][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833354 seconds
+[2025-12-21 12:14:32,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:10,132][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:10,252][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487606 seconds
+[2025-12-21 12:15:15,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490614 seconds
+[2025-12-21 12:15:15,257][openai._base_client][INFO] - Retrying request to /chat/completions in 0.437917 seconds
+[2025-12-21 12:15:15,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402658 seconds
+[2025-12-21 12:15:15,258][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447039 seconds
+[2025-12-21 12:15:15,713][openai._base_client][INFO] - Retrying request to /chat/completions in 0.855938 seconds
+[2025-12-21 12:15:19,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:22,068][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:24,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:24,973][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:25,119][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:25,326][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:25,810][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:28,267][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:29,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:34,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:39,449][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:39,566][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:41,927][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:47,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:47,868][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:48,844][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:48,874][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:50,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:52,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:53,446][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:58,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:15:59,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:01,751][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:12,208][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:14,805][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:18,534][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:21,022][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:39,708][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:40,578][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:45,587][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438255 seconds
+[2025-12-21 12:16:46,037][openai._base_client][INFO] - Retrying request to /chat/completions in 0.971100 seconds
+[2025-12-21 12:16:59,046][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:16:59,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468523 seconds
+[2025-12-21 12:17:04,233][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422048 seconds
+[2025-12-21 12:17:04,234][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390131 seconds
+[2025-12-21 12:17:04,234][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480259 seconds
+[2025-12-21 12:17:04,235][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383447 seconds
+[2025-12-21 12:17:04,235][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482988 seconds
+[2025-12-21 12:17:05,466][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:06,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:07,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:09,721][openai._base_client][INFO] - Retrying request to /chat/completions in 0.989269 seconds
+[2025-12-21 12:17:11,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:12,351][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:13,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:15,826][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:16,256][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:16,623][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:17,659][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:18,409][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:20,722][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436636 seconds
+[2025-12-21 12:17:25,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:26,164][openai._base_client][INFO] - Retrying request to /chat/completions in 0.965728 seconds
+[2025-12-21 12:17:27,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:27,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:27,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:29,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:30,673][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:31,534][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:33,712][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:36,304][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:39,356][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:39,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:17:43,564][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:04,551][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:06,112][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:15,692][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:16,027][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:20,491][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:25,503][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472591 seconds
+[2025-12-21 12:18:30,984][openai._base_client][INFO] - Retrying request to /chat/completions in 0.982469 seconds
+[2025-12-21 12:18:31,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:34,852][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:35,022][openai._base_client][INFO] - Retrying request to /chat/completions in 0.454475 seconds
+[2025-12-21 12:18:35,144][openai._base_client][INFO] - Retrying request to /chat/completions in 0.425134 seconds
+[2025-12-21 12:18:35,146][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461893 seconds
+[2025-12-21 12:18:35,153][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382219 seconds
+[2025-12-21 12:18:35,700][openai._base_client][INFO] - Retrying request to /chat/completions in 0.910632 seconds
+[2025-12-21 12:18:35,729][openai._base_client][INFO] - Retrying request to /chat/completions in 0.894940 seconds
+[2025-12-21 12:18:38,320][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:38,362][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:39,339][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:40,025][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416836 seconds
+[2025-12-21 12:18:40,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470414 seconds
+[2025-12-21 12:18:40,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432471 seconds
+[2025-12-21 12:18:40,597][openai._base_client][INFO] - Retrying request to /chat/completions in 0.756858 seconds
+[2025-12-21 12:18:40,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:40,686][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:41,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:43,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:43,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:43,702][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:43,874][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:47,031][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:47,076][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:47,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:48,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:51,151][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:51,550][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:51,942][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:51,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:54,353][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:54,853][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:57,452][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:18:57,919][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:19:00,115][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:19:01,192][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:19:06,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:19:20,979][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:00,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:05,391][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464980 seconds
+[2025-12-21 12:20:09,729][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:09,975][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459928 seconds
+[2025-12-21 12:20:09,993][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421614 seconds
+[2025-12-21 12:20:10,446][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863591 seconds
+[2025-12-21 12:20:13,061][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:13,255][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:13,494][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:14,357][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:14,592][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:14,962][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:16,209][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:16,724][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:17,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:17,162][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:17,488][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:17,729][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:17,856][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:18,684][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:19,460][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:19,987][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:21,772][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:22,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:22,669][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:22,900][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:24,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:24,694][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:29,534][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:34,130][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:37,413][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:39,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:20:44,950][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:29,639][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:29,956][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464085 seconds
+[2025-12-21 12:21:29,967][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424532 seconds
+[2025-12-21 12:21:29,974][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483368 seconds
+[2025-12-21 12:21:30,043][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416624 seconds
+[2025-12-21 12:21:32,912][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:34,021][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:34,202][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:34,228][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:34,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:34,923][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:34,959][openai._base_client][INFO] - Retrying request to /chat/completions in 0.423811 seconds
+[2025-12-21 12:21:34,960][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395494 seconds
+[2025-12-21 12:21:35,394][openai._base_client][INFO] - Retrying request to /chat/completions in 0.815656 seconds
+[2025-12-21 12:21:35,724][openai._base_client][INFO] - Retrying request to /chat/completions in 0.965989 seconds
+[2025-12-21 12:21:35,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:36,557][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:36,744][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:37,548][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:37,969][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:37,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:38,879][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:39,292][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:39,403][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:40,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:40,813][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:41,219][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:42,123][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:42,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:47,176][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:48,188][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:50,349][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:57,017][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:21:57,508][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:22:29,228][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:30:43,035][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/12-53-38/.hydra/config.yaml b/outputs/2025-12-21/12-53-38/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..085bbb56b0571501564f31a8da676665e16746e6
--- /dev/null
+++ b/outputs/2025-12-21/12-53-38/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - SimpleSudoku
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 64000
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/12-53-38/.hydra/hydra.yaml b/outputs/2025-12-21/12-53-38/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0296a91aebafe7c5e49a34d97be1b2a9ef15090f
--- /dev/null
+++ b/outputs/2025-12-21/12-53-38/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/12-53-38
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/12-53-38/.hydra/overrides.yaml b/outputs/2025-12-21/12-53-38/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/12-53-38/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/12-53-38/eval_api.log b/outputs/2025-12-21/12-53-38/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..1d4e4344c12abc68691009a6ccc1dc5f44e71042
--- /dev/null
+++ b/outputs/2025-12-21/12-53-38/eval_api.log
@@ -0,0 +1,355 @@
+[2025-12-21 12:53:44,932][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384873 seconds
+[2025-12-21 12:53:44,933][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428756 seconds
+[2025-12-21 12:53:44,934][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448079 seconds
+[2025-12-21 12:53:44,934][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462391 seconds
+[2025-12-21 12:53:44,934][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471947 seconds
+[2025-12-21 12:53:45,385][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:46,859][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:49,037][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:49,286][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:50,386][openai._base_client][INFO] - Retrying request to /chat/completions in 0.981116 seconds
+[2025-12-21 12:53:50,398][openai._base_client][INFO] - Retrying request to /chat/completions in 0.782392 seconds
+[2025-12-21 12:53:50,856][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:51,294][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:51,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:51,841][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:52,586][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:53,953][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:53,983][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:56,171][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:56,461][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:56,789][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:57,082][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:57,247][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:58,199][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:53:58,518][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:01,069][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:02,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:02,392][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:02,461][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:03,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:05,043][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:05,315][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:05,338][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:05,901][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:07,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:07,635][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:08,549][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:09,410][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:14,419][openai._base_client][INFO] - Retrying request to /chat/completions in 0.486348 seconds
+[2025-12-21 12:54:28,020][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:28,116][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471314 seconds
+[2025-12-21 12:54:32,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:33,117][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462270 seconds
+[2025-12-21 12:54:33,117][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424772 seconds
+[2025-12-21 12:54:33,118][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401127 seconds
+[2025-12-21 12:54:33,118][openai._base_client][INFO] - Retrying request to /chat/completions in 0.432195 seconds
+[2025-12-21 12:54:33,119][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470705 seconds
+[2025-12-21 12:54:33,376][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:33,799][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:33,867][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:34,926][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:35,115][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:35,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:36,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:37,547][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:38,523][openai._base_client][INFO] - Retrying request to /chat/completions in 0.788694 seconds
+[2025-12-21 12:54:38,584][openai._base_client][INFO] - Retrying request to /chat/completions in 0.762851 seconds
+[2025-12-21 12:54:38,733][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:39,679][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:39,886][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:40,774][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:42,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:42,608][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:42,678][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:43,140][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:43,483][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:43,964][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:44,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:44,334][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:44,900][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:45,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:46,382][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:46,557][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:47,461][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:48,190][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:48,464][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:51,694][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:53,077][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:53,854][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:54:58,863][openai._base_client][INFO] - Retrying request to /chat/completions in 0.433338 seconds
+[2025-12-21 12:55:05,581][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:05,735][openai._base_client][INFO] - Retrying request to /chat/completions in 0.484584 seconds
+[2025-12-21 12:55:10,835][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:11,102][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:11,184][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:11,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:11,402][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:11,642][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:11,810][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:12,440][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:12,463][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:12,512][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:12,792][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:14,369][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:15,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:15,625][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:15,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:15,887][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:16,122][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:16,801][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:17,057][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:17,228][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:18,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:18,031][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:18,220][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:18,671][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:18,672][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:21,389][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:21,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:21,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:21,990][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:23,447][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:23,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:28,420][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:28,545][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483694 seconds
+[2025-12-21 12:55:28,571][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387065 seconds
+[2025-12-21 12:55:28,573][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468017 seconds
+[2025-12-21 12:55:31,798][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:33,108][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:33,300][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:33,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:33,557][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:33,576][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377455 seconds
+[2025-12-21 12:55:33,576][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408595 seconds
+[2025-12-21 12:55:33,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406400 seconds
+[2025-12-21 12:55:33,577][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389233 seconds
+[2025-12-21 12:55:33,960][openai._base_client][INFO] - Retrying request to /chat/completions in 0.968739 seconds
+[2025-12-21 12:55:34,032][openai._base_client][INFO] - Retrying request to /chat/completions in 0.993268 seconds
+[2025-12-21 12:55:35,309][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:35,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:35,595][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:36,059][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:36,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:37,211][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:37,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:38,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:38,348][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:38,720][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:39,754][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:40,602][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:40,749][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:41,120][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:41,313][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:41,780][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:41,819][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:41,939][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:42,442][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:43,030][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:43,567][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:43,662][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:44,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:44,264][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:44,685][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:50,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:55:55,590][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375637 seconds
+[2025-12-21 12:56:04,085][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:04,273][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449745 seconds
+[2025-12-21 12:56:07,316][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:07,623][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:07,717][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:07,759][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:07,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:07,857][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:08,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:09,274][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388549 seconds
+[2025-12-21 12:56:09,275][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462844 seconds
+[2025-12-21 12:56:09,276][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385837 seconds
+[2025-12-21 12:56:09,276][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383779 seconds
+[2025-12-21 12:56:09,278][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441898 seconds
+[2025-12-21 12:56:09,528][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:09,725][openai._base_client][INFO] - Retrying request to /chat/completions in 0.994571 seconds
+[2025-12-21 12:56:10,155][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:10,473][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:10,514][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:10,929][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:10,950][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:10,961][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:11,208][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:11,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:11,802][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:12,425][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:13,007][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:13,454][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:13,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:13,895][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:14,164][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:14,537][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:14,539][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:14,663][openai._base_client][INFO] - Retrying request to /chat/completions in 0.886581 seconds
+[2025-12-21 12:56:14,664][openai._base_client][INFO] - Retrying request to /chat/completions in 0.859403 seconds
+[2025-12-21 12:56:14,665][openai._base_client][INFO] - Retrying request to /chat/completions in 0.904037 seconds
+[2025-12-21 12:56:14,741][openai._base_client][INFO] - Retrying request to /chat/completions in 0.963632 seconds
+[2025-12-21 12:56:14,774][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:15,773][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:18,071][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:18,573][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:19,233][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:20,042][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:25,052][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493232 seconds
+[2025-12-21 12:56:30,553][openai._base_client][INFO] - Retrying request to /chat/completions in 0.837734 seconds
+[2025-12-21 12:56:44,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:44,768][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474532 seconds
+[2025-12-21 12:56:47,991][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:48,302][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:48,944][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:49,165][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:49,401][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:49,405][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:49,768][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475942 seconds
+[2025-12-21 12:56:49,769][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476039 seconds
+[2025-12-21 12:56:49,770][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387413 seconds
+[2025-12-21 12:56:49,770][openai._base_client][INFO] - Retrying request to /chat/completions in 0.375470 seconds
+[2025-12-21 12:56:49,771][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413177 seconds
+[2025-12-21 12:56:49,771][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489226 seconds
+[2025-12-21 12:56:49,771][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420149 seconds
+[2025-12-21 12:56:49,772][openai._base_client][INFO] - Retrying request to /chat/completions in 0.471633 seconds
+[2025-12-21 12:56:50,138][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:50,245][openai._base_client][INFO] - Retrying request to /chat/completions in 0.851825 seconds
+[2025-12-21 12:56:51,116][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:51,226][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:51,445][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:52,645][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:52,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:53,031][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:53,354][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:54,441][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:54,866][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:54,935][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:55,148][openai._base_client][INFO] - Retrying request to /chat/completions in 0.784936 seconds
+[2025-12-21 12:56:55,161][openai._base_client][INFO] - Retrying request to /chat/completions in 0.928043 seconds
+[2025-12-21 12:56:55,186][openai._base_client][INFO] - Retrying request to /chat/completions in 0.787229 seconds
+[2025-12-21 12:56:55,194][openai._base_client][INFO] - Retrying request to /chat/completions in 0.839167 seconds
+[2025-12-21 12:56:55,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:55,246][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967494 seconds
+[2025-12-21 12:56:55,264][openai._base_client][INFO] - Retrying request to /chat/completions in 0.997775 seconds
+[2025-12-21 12:56:55,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:55,901][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:57,234][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:57,293][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:57,344][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:57,482][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:58,496][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:58,644][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:59,508][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:59,706][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:56:59,896][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:00,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:06,274][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383128 seconds
+[2025-12-21 12:57:11,282][openai._base_client][INFO] - Retrying request to /chat/completions in 0.390260 seconds
+[2025-12-21 12:57:11,660][openai._base_client][INFO] - Retrying request to /chat/completions in 0.858834 seconds
+[2025-12-21 12:57:16,670][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:22,532][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442490 seconds
+[2025-12-21 12:57:28,247][openai._base_client][INFO] - Retrying request to /chat/completions in 0.864427 seconds
+[2025-12-21 12:57:33,553][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:33,674][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431941 seconds
+[2025-12-21 12:57:37,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:38,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:38,674][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462964 seconds
+[2025-12-21 12:57:38,675][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463235 seconds
+[2025-12-21 12:57:38,675][openai._base_client][INFO] - Retrying request to /chat/completions in 0.436365 seconds
+[2025-12-21 12:57:38,676][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384018 seconds
+[2025-12-21 12:57:38,676][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408529 seconds
+[2025-12-21 12:57:38,677][openai._base_client][INFO] - Retrying request to /chat/completions in 0.479529 seconds
+[2025-12-21 12:57:38,677][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400299 seconds
+[2025-12-21 12:57:39,109][openai._base_client][INFO] - Retrying request to /chat/completions in 0.772199 seconds
+[2025-12-21 12:57:42,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:42,480][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:43,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:44,089][openai._base_client][INFO] - Retrying request to /chat/completions in 0.844236 seconds
+[2025-12-21 12:57:44,115][openai._base_client][INFO] - Retrying request to /chat/completions in 0.973585 seconds
+[2025-12-21 12:57:44,140][openai._base_client][INFO] - Retrying request to /chat/completions in 0.811261 seconds
+[2025-12-21 12:57:44,141][openai._base_client][INFO] - Retrying request to /chat/completions in 0.804590 seconds
+[2025-12-21 12:57:45,308][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:47,607][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:47,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:48,364][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:57:55,101][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446093 seconds
+[2025-12-21 12:58:00,452][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:00,507][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419588 seconds
+[2025-12-21 12:58:04,621][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:05,191][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:05,506][openai._base_client][INFO] - Retrying request to /chat/completions in 0.450805 seconds
+[2025-12-21 12:58:05,507][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383878 seconds
+[2025-12-21 12:58:09,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:10,574][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:10,610][openai._base_client][INFO] - Retrying request to /chat/completions in 0.431637 seconds
+[2025-12-21 12:58:10,611][openai._base_client][INFO] - Retrying request to /chat/completions in 0.389775 seconds
+[2025-12-21 12:58:14,830][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:16,088][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:16,119][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403758 seconds
+[2025-12-21 12:58:16,525][openai._base_client][INFO] - Retrying request to /chat/completions in 0.898113 seconds
+[2025-12-21 12:58:22,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:22,855][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422090 seconds
+[2025-12-21 12:58:27,907][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:27,938][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474195 seconds
+[2025-12-21 12:58:32,767][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:32,816][openai._base_client][INFO] - Retrying request to /chat/completions in 0.393532 seconds
+[2025-12-21 12:58:38,216][openai._base_client][INFO] - Retrying request to /chat/completions in 0.751384 seconds
+[2025-12-21 12:58:53,065][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:53,094][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457217 seconds
+[2025-12-21 12:58:59,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:58:59,587][openai._base_client][INFO] - Retrying request to /chat/completions in 0.492297 seconds
+[2025-12-21 12:59:05,087][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916285 seconds
+[2025-12-21 12:59:10,185][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:59:10,216][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404301 seconds
+[2025-12-21 12:59:15,318][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:59:15,344][root][WARNING] - Truncated history for env 28 to keep prompt within 7200 tokens (current 7161).
+[2025-12-21 12:59:15,369][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385047 seconds
+[2025-12-21 12:59:35,683][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:59:35,707][root][WARNING] - Truncated history for env 28 to keep prompt within 7200 tokens (current 7007).
+[2025-12-21 12:59:35,727][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470882 seconds
+[2025-12-21 12:59:41,206][openai._base_client][INFO] - Retrying request to /chat/completions in 0.916006 seconds
+[2025-12-21 12:59:49,705][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 12:59:49,731][root][WARNING] - Truncated history for env 28 to keep prompt within 7200 tokens (current 6903).
+[2025-12-21 12:59:49,749][openai._base_client][INFO] - Retrying request to /chat/completions in 0.465015 seconds
+[2025-12-21 13:00:00,443][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:00:00,478][root][WARNING] - Truncated history for env 28 to keep prompt within 7200 tokens (current 6794).
+[2025-12-21 13:00:00,496][openai._base_client][INFO] - Retrying request to /chat/completions in 0.451520 seconds
+[2025-12-21 13:00:05,955][openai._base_client][INFO] - Retrying request to /chat/completions in 0.798328 seconds
+[2025-12-21 13:00:18,661][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:00:18,664][root][WARNING] - Skipping non-numeric metric 'error' with value 'Number 1 conflicts with column 2' for env SimpleSudoku.
+[2025-12-21 13:00:18,664][root][WARNING] - Skipping non-numeric metric 'valid_numbers' with value [3] for env SimpleSudoku.
+[2025-12-21 13:00:18,665][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,665][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,665][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,665][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,666][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,666][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,667][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,667][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,667][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,668][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,668][root][WARNING] - Skipping non-numeric metric 'error' with value 'Number 2 conflicts with row 3, column 2, box (2,1)' for env SimpleSudoku.
+[2025-12-21 13:00:18,668][root][WARNING] - Skipping non-numeric metric 'valid_numbers' with value [] for env SimpleSudoku.
+[2025-12-21 13:00:18,668][root][WARNING] - Skipping non-numeric metric 'error' with value 'Number 1 conflicts with row 1, box (1,2)' for env SimpleSudoku.
+[2025-12-21 13:00:18,668][root][WARNING] - Skipping non-numeric metric 'valid_numbers' with value [4] for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Number 4 conflicts with row 4' for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'valid_numbers' with value [1] for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,669][root][WARNING] - Skipping non-numeric metric 'error' with value 'Number 1 conflicts with row 1, box (1,2)' for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'valid_numbers' with value [4] for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'error' with value 'Number 4 conflicts with row 4' for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'valid_numbers' with value [1] for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'error' with value 'Number 1 conflicts with row 1, box (1,2)' for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'valid_numbers' with value [4] for env SimpleSudoku.
+[2025-12-21 13:00:18,670][root][WARNING] - Skipping non-numeric metric 'error' with value 'Cannot modify initial cells' for env SimpleSudoku.
diff --git a/outputs/2025-12-21/13-10-04/.hydra/config.yaml b/outputs/2025-12-21/13-10-04/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ab4f642e6f9320059cfbed52b710b035f1944169
--- /dev/null
+++ b/outputs/2025-12-21/13-10-04/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 64000
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/13-10-04/.hydra/hydra.yaml b/outputs/2025-12-21/13-10-04/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..c44d343ce1815727168addcee87cddf125cf79c9
--- /dev/null
+++ b/outputs/2025-12-21/13-10-04/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/13-10-04
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/13-10-04/.hydra/overrides.yaml b/outputs/2025-12-21/13-10-04/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/13-10-04/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/13-10-04/eval_api.log b/outputs/2025-12-21/13-10-04/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..3d08f08052ae8d27e1012b8ea53c0f734643dff3
--- /dev/null
+++ b/outputs/2025-12-21/13-10-04/eval_api.log
@@ -0,0 +1,182 @@
+[2025-12-21 13:10:11,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:11,937][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:11,980][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:12,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:12,608][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:12,955][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:13,340][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:13,456][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:13,478][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:13,524][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:13,641][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:13,848][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:13,880][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:14,988][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:15,216][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:17,252][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:17,471][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:19,343][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:19,366][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:19,601][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:20,390][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:21,610][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:21,933][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:22,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:22,929][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:22,981][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:23,098][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:23,546][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:24,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:25,291][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:29,627][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:33,476][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:33,559][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434928 seconds
+[2025-12-21 13:10:33,560][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466539 seconds
+[2025-12-21 13:10:33,662][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408095 seconds
+[2025-12-21 13:10:33,669][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397126 seconds
+[2025-12-21 13:10:33,671][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388888 seconds
+[2025-12-21 13:10:33,672][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472049 seconds
+[2025-12-21 13:10:37,513][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:38,154][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:38,203][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:38,224][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:38,562][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377281 seconds
+[2025-12-21 13:10:38,562][openai._base_client][INFO] - Retrying request to /chat/completions in 0.416461 seconds
+[2025-12-21 13:10:38,563][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476449 seconds
+[2025-12-21 13:10:38,563][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440986 seconds
+[2025-12-21 13:10:38,564][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474114 seconds
+[2025-12-21 13:10:38,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:38,997][openai._base_client][INFO] - Retrying request to /chat/completions in 0.793215 seconds
+[2025-12-21 13:10:39,030][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869851 seconds
+[2025-12-21 13:10:39,063][openai._base_client][INFO] - Retrying request to /chat/completions in 0.960206 seconds
+[2025-12-21 13:10:39,075][openai._base_client][INFO] - Retrying request to /chat/completions in 0.895078 seconds
+[2025-12-21 13:10:39,520][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:40,993][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:41,766][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:41,790][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:41,985][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:42,238][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:43,086][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:43,875][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:43,942][openai._base_client][INFO] - Retrying request to /chat/completions in 0.987928 seconds
+[2025-12-21 13:10:43,983][openai._base_client][INFO] - Retrying request to /chat/completions in 0.784907 seconds
+[2025-12-21 13:10:43,996][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:44,042][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999417 seconds
+[2025-12-21 13:10:44,396][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:45,242][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:45,273][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:45,823][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:45,891][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:47,018][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:47,105][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:47,865][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:47,944][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:50,034][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:50,116][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:50,127][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:10:50,261][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:03,556][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:08,566][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466724 seconds
+[2025-12-21 13:11:12,720][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:14,039][openai._base_client][INFO] - Retrying request to /chat/completions in 0.801105 seconds
+[2025-12-21 13:11:18,871][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:18,961][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469935 seconds
+[2025-12-21 13:11:22,838][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:22,919][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,012][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,067][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,103][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,305][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,374][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,492][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,808][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:23,962][openai._base_client][INFO] - Retrying request to /chat/completions in 0.499357 seconds
+[2025-12-21 13:11:23,963][openai._base_client][INFO] - Retrying request to /chat/completions in 0.494740 seconds
+[2025-12-21 13:11:23,965][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395714 seconds
+[2025-12-21 13:11:23,966][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421977 seconds
+[2025-12-21 13:11:23,966][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447550 seconds
+[2025-12-21 13:11:23,989][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:25,054][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:25,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:26,342][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:26,701][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:26,886][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:26,889][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:27,210][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:27,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:27,478][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:27,761][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:28,386][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:28,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:29,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:29,189][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413220 seconds
+[2025-12-21 13:11:29,191][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406712 seconds
+[2025-12-21 13:11:29,193][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476333 seconds
+[2025-12-21 13:11:29,195][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377829 seconds
+[2025-12-21 13:11:29,196][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439593 seconds
+[2025-12-21 13:11:29,197][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403219 seconds
+[2025-12-21 13:11:29,198][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408098 seconds
+[2025-12-21 13:11:29,199][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427319 seconds
+[2025-12-21 13:11:29,201][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417755 seconds
+[2025-12-21 13:11:29,202][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480388 seconds
+[2025-12-21 13:11:29,202][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418578 seconds
+[2025-12-21 13:11:29,203][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424886 seconds
+[2025-12-21 13:11:29,204][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466601 seconds
+[2025-12-21 13:11:33,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:33,368][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:33,479][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:33,501][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:33,612][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:33,766][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:34,096][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:34,169][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:34,205][openai._base_client][INFO] - Retrying request to /chat/completions in 0.403466 seconds
+[2025-12-21 13:11:34,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:34,603][openai._base_client][INFO] - Retrying request to /chat/completions in 0.927068 seconds
+[2025-12-21 13:11:34,606][openai._base_client][INFO] - Retrying request to /chat/completions in 0.858614 seconds
+[2025-12-21 13:11:34,621][openai._base_client][INFO] - Retrying request to /chat/completions in 0.984587 seconds
+[2025-12-21 13:11:34,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.794493 seconds
+[2025-12-21 13:11:34,630][openai._base_client][INFO] - Retrying request to /chat/completions in 0.822918 seconds
+[2025-12-21 13:11:34,684][openai._base_client][INFO] - Retrying request to /chat/completions in 0.857016 seconds
+[2025-12-21 13:11:35,341][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:37,031][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:37,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:38,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:39,078][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:39,145][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:39,506][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:39,689][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:39,749][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402495 seconds
+[2025-12-21 13:11:39,750][openai._base_client][INFO] - Retrying request to /chat/completions in 0.474724 seconds
+[2025-12-21 13:11:39,751][openai._base_client][INFO] - Retrying request to /chat/completions in 0.430456 seconds
+[2025-12-21 13:11:39,752][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442597 seconds
+[2025-12-21 13:11:39,753][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447568 seconds
+[2025-12-21 13:11:39,754][openai._base_client][INFO] - Retrying request to /chat/completions in 0.487193 seconds
+[2025-12-21 13:11:39,755][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452482 seconds
+[2025-12-21 13:11:42,599][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:43,839][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:44,244][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:44,755][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495929 seconds
+[2025-12-21 13:11:45,154][openai._base_client][INFO] - Retrying request to /chat/completions in 0.857447 seconds
+[2025-12-21 13:11:45,184][openai._base_client][INFO] - Retrying request to /chat/completions in 0.810869 seconds
+[2025-12-21 13:11:45,197][openai._base_client][INFO] - Retrying request to /chat/completions in 0.795825 seconds
+[2025-12-21 13:11:45,210][openai._base_client][INFO] - Retrying request to /chat/completions in 0.752286 seconds
+[2025-12-21 13:11:45,226][openai._base_client][INFO] - Retrying request to /chat/completions in 0.871876 seconds
+[2025-12-21 13:11:48,269][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:48,580][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:49,008][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:49,774][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:49,908][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:50,511][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:11:50,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409707 seconds
+[2025-12-21 13:11:50,534][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435453 seconds
+[2025-12-21 13:11:50,947][openai._base_client][INFO] - Retrying request to /chat/completions in 0.763243 seconds
+[2025-12-21 13:11:50,971][openai._base_client][INFO] - Retrying request to /chat/completions in 0.872955 seconds
+[2025-12-21 13:11:59,698][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:12:01,393][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:12:01,408][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439430 seconds
+[2025-12-21 13:12:01,850][openai._base_client][INFO] - Retrying request to /chat/completions in 0.955298 seconds
+[2025-12-21 13:12:15,515][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:12:15,540][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385795 seconds
+[2025-12-21 13:12:19,664][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/13-13-01/.hydra/config.yaml b/outputs/2025-12-21/13-13-01/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..ab4f642e6f9320059cfbed52b710b035f1944169
--- /dev/null
+++ b/outputs/2025-12-21/13-13-01/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 7200
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 25
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 32
+    group_size: 1
+    env_configs:
+      tags:
+      - CoordFrozenLake
+      n_groups:
+      - 32
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 64000
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/13-13-01/eval_api.log b/outputs/2025-12-21/13-13-01/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..6b8aa778d98c7447838100c0ba0e3247f497dc65
--- /dev/null
+++ b/outputs/2025-12-21/13-13-01/eval_api.log
@@ -0,0 +1,275 @@
+[2025-12-21 13:13:08,137][openai._base_client][INFO] - Retrying request to /chat/completions in 0.384222 seconds
+[2025-12-21 13:13:08,138][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377507 seconds
+[2025-12-21 13:13:08,138][openai._base_client][INFO] - Retrying request to /chat/completions in 0.413004 seconds
+[2025-12-21 13:13:08,139][openai._base_client][INFO] - Retrying request to /chat/completions in 0.473551 seconds
+[2025-12-21 13:13:08,787][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:08,871][openai._base_client][INFO] - Retrying request to /chat/completions in 0.897032 seconds
+[2025-12-21 13:13:09,213][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:09,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:10,962][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:11,164][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:11,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:11,531][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:11,555][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:12,050][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:12,322][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:12,766][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:13,519][openai._base_client][INFO] - Retrying request to /chat/completions in 0.930080 seconds
+[2025-12-21 13:13:13,535][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:13,554][openai._base_client][INFO] - Retrying request to /chat/completions in 0.753658 seconds
+[2025-12-21 13:13:13,615][openai._base_client][INFO] - Retrying request to /chat/completions in 0.941784 seconds
+[2025-12-21 13:13:14,888][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:15,136][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498140 seconds
+[2025-12-21 13:13:15,576][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:16,565][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:18,403][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:18,469][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:19,911][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417304 seconds
+[2025-12-21 13:13:20,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:20,336][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:20,637][openai._base_client][INFO] - Retrying request to /chat/completions in 0.773951 seconds
+[2025-12-21 13:13:20,730][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:20,752][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:21,209][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:21,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:23,834][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:24,332][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:25,121][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:28,100][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:38,225][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:43,235][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448482 seconds
+[2025-12-21 13:13:43,588][openai._base_client][INFO] - Retrying request to /chat/completions in 0.426302 seconds
+[2025-12-21 13:13:43,600][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401059 seconds
+[2025-12-21 13:13:44,348][openai._base_client][INFO] - Retrying request to /chat/completions in 0.999251 seconds
+[2025-12-21 13:13:48,243][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379159 seconds
+[2025-12-21 13:13:52,699][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:52,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:13:53,627][openai._base_client][INFO] - Retrying request to /chat/completions in 0.959864 seconds
+[2025-12-21 13:14:00,643][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:05,652][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453947 seconds
+[2025-12-21 13:14:15,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:15,599][openai._base_client][INFO] - Retrying request to /chat/completions in 0.446600 seconds
+[2025-12-21 13:14:15,948][openai._base_client][INFO] - Retrying request to /chat/completions in 0.441449 seconds
+[2025-12-21 13:14:15,959][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406626 seconds
+[2025-12-21 13:14:15,961][openai._base_client][INFO] - Retrying request to /chat/completions in 0.401258 seconds
+[2025-12-21 13:14:15,966][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488133 seconds
+[2025-12-21 13:14:15,967][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498227 seconds
+[2025-12-21 13:14:15,969][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417830 seconds
+[2025-12-21 13:14:16,707][openai._base_client][INFO] - Retrying request to /chat/completions in 0.811252 seconds
+[2025-12-21 13:14:16,739][openai._base_client][INFO] - Retrying request to /chat/completions in 0.752315 seconds
+[2025-12-21 13:14:20,423][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:20,603][openai._base_client][INFO] - Retrying request to /chat/completions in 0.417313 seconds
+[2025-12-21 13:14:20,603][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387616 seconds
+[2025-12-21 13:14:20,604][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383244 seconds
+[2025-12-21 13:14:20,604][openai._base_client][INFO] - Retrying request to /chat/completions in 0.405297 seconds
+[2025-12-21 13:14:20,605][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434621 seconds
+[2025-12-21 13:14:20,605][openai._base_client][INFO] - Retrying request to /chat/completions in 0.396114 seconds
+[2025-12-21 13:14:20,618][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:20,849][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:21,049][openai._base_client][INFO] - Retrying request to /chat/completions in 0.896968 seconds
+[2025-12-21 13:14:21,366][openai._base_client][INFO] - Retrying request to /chat/completions in 0.824382 seconds
+[2025-12-21 13:14:21,390][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967669 seconds
+[2025-12-21 13:14:21,457][openai._base_client][INFO] - Retrying request to /chat/completions in 0.944994 seconds
+[2025-12-21 13:14:21,467][openai._base_client][INFO] - Retrying request to /chat/completions in 0.875159 seconds
+[2025-12-21 13:14:22,853][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445422 seconds
+[2025-12-21 13:14:25,378][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:25,991][openai._base_client][INFO] - Retrying request to /chat/completions in 0.932102 seconds
+[2025-12-21 13:14:25,994][openai._base_client][INFO] - Retrying request to /chat/completions in 0.937149 seconds
+[2025-12-21 13:14:26,004][openai._base_client][INFO] - Retrying request to /chat/completions in 0.990434 seconds
+[2025-12-21 13:14:26,012][openai._base_client][INFO] - Retrying request to /chat/completions in 0.962504 seconds
+[2025-12-21 13:14:26,024][openai._base_client][INFO] - Retrying request to /chat/completions in 0.846691 seconds
+[2025-12-21 13:14:26,041][openai._base_client][INFO] - Retrying request to /chat/completions in 0.948256 seconds
+[2025-12-21 13:14:26,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:27,301][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472793 seconds
+[2025-12-21 13:14:27,524][openai._base_client][INFO] - Retrying request to /chat/completions in 0.386323 seconds
+[2025-12-21 13:14:27,560][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445479 seconds
+[2025-12-21 13:14:27,646][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376317 seconds
+[2025-12-21 13:14:28,271][openai._base_client][INFO] - Retrying request to /chat/completions in 0.820788 seconds
+[2025-12-21 13:14:28,302][openai._base_client][INFO] - Retrying request to /chat/completions in 0.994015 seconds
+[2025-12-21 13:14:28,373][openai._base_client][INFO] - Retrying request to /chat/completions in 0.768739 seconds
+[2025-12-21 13:14:29,144][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:29,431][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:29,530][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:32,349][openai._base_client][INFO] - Retrying request to /chat/completions in 0.427894 seconds
+[2025-12-21 13:14:32,358][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498296 seconds
+[2025-12-21 13:14:32,362][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463425 seconds
+[2025-12-21 13:14:32,778][openai._base_client][INFO] - Retrying request to /chat/completions in 0.863534 seconds
+[2025-12-21 13:14:33,010][openai._base_client][INFO] - Retrying request to /chat/completions in 0.869612 seconds
+[2025-12-21 13:14:33,814][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:37,035][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:37,523][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:37,929][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:37,949][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:39,270][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:44,279][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391000 seconds
+[2025-12-21 13:14:44,633][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409209 seconds
+[2025-12-21 13:14:44,634][openai._base_client][INFO] - Retrying request to /chat/completions in 0.462943 seconds
+[2025-12-21 13:14:44,648][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456834 seconds
+[2025-12-21 13:14:44,650][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468497 seconds
+[2025-12-21 13:14:44,663][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472321 seconds
+[2025-12-21 13:14:44,670][openai._base_client][INFO] - Retrying request to /chat/completions in 0.412614 seconds
+[2025-12-21 13:14:45,457][openai._base_client][INFO] - Retrying request to /chat/completions in 0.832156 seconds
+[2025-12-21 13:14:48,647][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:48,994][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:49,303][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409126 seconds
+[2025-12-21 13:14:49,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.482821 seconds
+[2025-12-21 13:14:49,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443269 seconds
+[2025-12-21 13:14:49,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407066 seconds
+[2025-12-21 13:14:49,305][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378500 seconds
+[2025-12-21 13:14:49,674][openai._base_client][INFO] - Retrying request to /chat/completions in 0.824800 seconds
+[2025-12-21 13:14:50,110][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:50,323][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:50,832][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:50,951][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:52,699][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:54,259][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:54,532][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:54,718][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:54,750][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967321 seconds
+[2025-12-21 13:14:55,193][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:55,331][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:14:55,596][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:00,958][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:05,966][openai._base_client][INFO] - Retrying request to /chat/completions in 0.480557 seconds
+[2025-12-21 13:15:06,777][openai._base_client][INFO] - Retrying request to /chat/completions in 0.804789 seconds
+[2025-12-21 13:15:22,601][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391532 seconds
+[2025-12-21 13:15:26,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:27,068][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460359 seconds
+[2025-12-21 13:15:30,111][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:31,237][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:31,639][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:31,720][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:31,983][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:32,069][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439906 seconds
+[2025-12-21 13:15:32,070][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376227 seconds
+[2025-12-21 13:15:32,073][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418775 seconds
+[2025-12-21 13:15:32,073][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440334 seconds
+[2025-12-21 13:15:32,232][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:32,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:33,507][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:33,526][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:33,721][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:33,921][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:34,498][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:34,714][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:35,717][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:35,959][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:36,147][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:36,338][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:36,658][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:36,700][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:36,723][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:36,748][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:36,794][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:37,327][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:37,371][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:37,512][openai._base_client][INFO] - Retrying request to /chat/completions in 0.865844 seconds
+[2025-12-21 13:15:37,516][openai._base_client][INFO] - Retrying request to /chat/completions in 0.797894 seconds
+[2025-12-21 13:15:38,179][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:41,617][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:42,467][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:42,618][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435145 seconds
+[2025-12-21 13:15:42,619][openai._base_client][INFO] - Retrying request to /chat/completions in 0.493882 seconds
+[2025-12-21 13:15:42,623][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397995 seconds
+[2025-12-21 13:15:46,360][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:46,899][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:47,041][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:47,104][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:47,204][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:47,400][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:47,623][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381593 seconds
+[2025-12-21 13:15:47,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438436 seconds
+[2025-12-21 13:15:47,624][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495986 seconds
+[2025-12-21 13:15:47,651][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:48,023][openai._base_client][INFO] - Retrying request to /chat/completions in 0.767963 seconds
+[2025-12-21 13:15:48,097][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:48,115][openai._base_client][INFO] - Retrying request to /chat/completions in 0.852668 seconds
+[2025-12-21 13:15:48,484][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:49,066][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:49,954][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:50,056][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:50,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:50,486][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:50,638][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:50,905][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:51,148][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:51,966][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:52,092][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:52,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:52,797][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:53,010][openai._base_client][INFO] - Retrying request to /chat/completions in 0.830983 seconds
+[2025-12-21 13:15:53,422][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:57,046][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:15:57,120][openai._base_client][INFO] - Retrying request to /chat/completions in 0.421544 seconds
+[2025-12-21 13:15:57,150][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435963 seconds
+[2025-12-21 13:15:57,151][openai._base_client][INFO] - Retrying request to /chat/completions in 0.459117 seconds
+[2025-12-21 13:15:57,152][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408610 seconds
+[2025-12-21 13:15:57,154][openai._base_client][INFO] - Retrying request to /chat/completions in 0.383545 seconds
+[2025-12-21 13:16:00,688][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,101][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,205][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,355][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,438][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,464][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,821][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,933][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:01,978][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:02,541][openai._base_client][INFO] - Retrying request to /chat/completions in 0.809277 seconds
+[2025-12-21 13:16:02,614][openai._base_client][INFO] - Retrying request to /chat/completions in 0.776819 seconds
+[2025-12-21 13:16:02,894][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:03,038][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:03,825][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:03,904][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:06,598][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:07,626][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:13,319][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:13,427][openai._base_client][INFO] - Retrying request to /chat/completions in 0.463295 seconds
+[2025-12-21 13:16:17,093][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:17,272][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:17,357][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:17,380][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:17,829][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:18,430][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434231 seconds
+[2025-12-21 13:16:18,430][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464110 seconds
+[2025-12-21 13:16:18,431][openai._base_client][INFO] - Retrying request to /chat/completions in 0.407303 seconds
+[2025-12-21 13:16:18,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.388432 seconds
+[2025-12-21 13:16:18,893][openai._base_client][INFO] - Retrying request to /chat/completions in 0.807425 seconds
+[2025-12-21 13:16:19,321][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:19,361][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:20,750][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:21,246][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:22,498][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:22,605][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:22,840][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:33,294][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:33,370][openai._base_client][INFO] - Retrying request to /chat/completions in 0.461789 seconds
+[2025-12-21 13:16:36,297][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:37,398][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:37,515][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:37,536][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:38,330][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:38,340][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:38,346][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:38,370][openai._base_client][INFO] - Retrying request to /chat/completions in 0.439293 seconds
+[2025-12-21 13:16:38,371][openai._base_client][INFO] - Retrying request to /chat/completions in 0.378533 seconds
+[2025-12-21 13:16:41,603][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:43,178][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:43,230][openai._base_client][INFO] - Retrying request to /chat/completions in 0.449826 seconds
+[2025-12-21 13:16:43,231][openai._base_client][INFO] - Retrying request to /chat/completions in 0.448225 seconds
+[2025-12-21 13:16:43,232][openai._base_client][INFO] - Retrying request to /chat/completions in 0.376250 seconds
+[2025-12-21 13:16:43,233][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419013 seconds
+[2025-12-21 13:16:46,367][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:47,943][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:48,235][openai._base_client][INFO] - Retrying request to /chat/completions in 0.429029 seconds
+[2025-12-21 13:16:48,655][openai._base_client][INFO] - Retrying request to /chat/completions in 0.783053 seconds
+[2025-12-21 13:16:48,682][openai._base_client][INFO] - Retrying request to /chat/completions in 0.833220 seconds
+[2025-12-21 13:16:51,062][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:51,756][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:16:59,535][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452377 seconds
+[2025-12-21 13:17:04,994][openai._base_client][INFO] - Retrying request to /chat/completions in 0.856766 seconds
+[2025-12-21 13:17:20,874][openai._base_client][INFO] - Retrying request to /chat/completions in 0.453605 seconds
+[2025-12-21 13:17:23,982][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:17:24,010][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422314 seconds
+[2025-12-21 13:17:27,048][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:17:28,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
diff --git a/outputs/2025-12-21/13-21-47/.hydra/hydra.yaml b/outputs/2025-12-21/13-21-47/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..047cf6d8f557c504204dd3bc7e62dfb80198afff
--- /dev/null
+++ b/outputs/2025-12-21/13-21-47/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/13-21-47
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/13-27-06/.hydra/hydra.yaml b/outputs/2025-12-21/13-27-06/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..d0a44c278f08cf87d3f92d16154a7c5ffa0ec636
--- /dev/null
+++ b/outputs/2025-12-21/13-27-06/.hydra/hydra.yaml
@@ -0,0 +1,171 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath:
+  - file://./verl/verl/trainer/config
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.searchpath=[file://./verl/verl/trainer/config]
+    - hydra.mode=RUN
+    task: []
+  job:
+    name: eval_api
+    chdir: null
+    override_dirname: ''
+    id: ???
+    num: ???
+    config_name: evaluate_api_llm
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: ./verl/verl/trainer/config
+      schema: file
+      provider: hydra.searchpath in command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/13-27-06
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/13-27-06/.hydra/overrides.yaml b/outputs/2025-12-21/13-27-06/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fe51488c7066f6687ef680d6bfaa4f7768ef205c
--- /dev/null
+++ b/outputs/2025-12-21/13-27-06/.hydra/overrides.yaml
@@ -0,0 +1 @@
+[]
diff --git a/outputs/2025-12-21/13-27-06/eval_api.log b/outputs/2025-12-21/13-27-06/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..5c1264c9571a527973d33f9eece30918f1717b98
--- /dev/null
+++ b/outputs/2025-12-21/13-27-06/eval_api.log
@@ -0,0 +1,147 @@
+[2025-12-21 13:27:18,462][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:27:18,476][openai._base_client][INFO] - Retrying request to /chat/completions in 0.456852 seconds
+[2025-12-21 13:27:27,893][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:27:27,903][openai._base_client][INFO] - Retrying request to /chat/completions in 0.434463 seconds
+[2025-12-21 13:27:33,345][openai._base_client][INFO] - Retrying request to /chat/completions in 0.862433 seconds
+[2025-12-21 13:27:43,569][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:27:43,582][openai._base_client][INFO] - Retrying request to /chat/completions in 0.452985 seconds
+[2025-12-21 13:27:49,776][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:27:49,793][openai._base_client][INFO] - Retrying request to /chat/completions in 0.470523 seconds
+[2025-12-21 13:27:58,249][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:27:58,292][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402117 seconds
+[2025-12-21 13:28:03,704][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806730 seconds
+[2025-12-21 13:28:19,533][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467455 seconds
+[2025-12-21 13:28:25,007][openai._base_client][INFO] - Retrying request to /chat/completions in 0.764675 seconds
+[2025-12-21 13:28:31,608][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:28:31,628][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422289 seconds
+[2025-12-21 13:28:37,059][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985556 seconds
+[2025-12-21 13:28:53,067][openai._base_client][INFO] - Retrying request to /chat/completions in 0.395832 seconds
+[2025-12-21 13:29:00,886][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:29:00,936][openai._base_client][INFO] - Retrying request to /chat/completions in 0.442433 seconds
+[2025-12-21 13:29:08,411][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:29:08,438][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408577 seconds
+[2025-12-21 13:29:13,855][openai._base_client][INFO] - Retrying request to /chat/completions in 0.963478 seconds
+[2025-12-21 13:29:49,679][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:29:49,714][openai._base_client][INFO] - Retrying request to /chat/completions in 0.381243 seconds
+[2025-12-21 13:29:58,345][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:29:58,384][openai._base_client][INFO] - Retrying request to /chat/completions in 0.414364 seconds
+[2025-12-21 13:30:03,807][openai._base_client][INFO] - Retrying request to /chat/completions in 0.802915 seconds
+[2025-12-21 13:30:15,410][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:30:15,451][openai._base_client][INFO] - Retrying request to /chat/completions in 0.402718 seconds
+[2025-12-21 13:30:20,860][openai._base_client][INFO] - Retrying request to /chat/completions in 0.831603 seconds
+[2025-12-21 13:30:57,160][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:30:57,215][openai._base_client][INFO] - Retrying request to /chat/completions in 0.495389 seconds
+[2025-12-21 13:31:14,224][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:31:14,284][openai._base_client][INFO] - Retrying request to /chat/completions in 0.498955 seconds
+[2025-12-21 13:31:19,788][openai._base_client][INFO] - Retrying request to /chat/completions in 0.902045 seconds
+[2025-12-21 13:31:27,153][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:31:27,215][openai._base_client][INFO] - Retrying request to /chat/completions in 0.467927 seconds
+[2025-12-21 13:31:33,333][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:31:33,401][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424573 seconds
+[2025-12-21 13:31:38,833][openai._base_client][INFO] - Retrying request to /chat/completions in 0.830876 seconds
+[2025-12-21 13:31:54,015][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:31:54,097][openai._base_client][INFO] - Retrying request to /chat/completions in 0.419345 seconds
+[2025-12-21 13:31:59,524][openai._base_client][INFO] - Retrying request to /chat/completions in 0.861097 seconds
+[2025-12-21 13:32:11,352][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:32:11,437][openai._base_client][INFO] - Retrying request to /chat/completions in 0.379358 seconds
+[2025-12-21 13:32:16,824][openai._base_client][INFO] - Retrying request to /chat/completions in 0.806030 seconds
+[2025-12-21 13:32:37,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:32:37,619][openai._base_client][INFO] - Retrying request to /chat/completions in 0.377548 seconds
+[2025-12-21 13:32:43,004][openai._base_client][INFO] - Retrying request to /chat/completions in 0.952018 seconds
+[2025-12-21 13:32:58,982][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444344 seconds
+[2025-12-21 13:33:04,436][openai._base_client][INFO] - Retrying request to /chat/completions in 0.903672 seconds
+[2025-12-21 13:33:14,956][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:33:15,135][openai._base_client][INFO] - Retrying request to /chat/completions in 0.385633 seconds
+[2025-12-21 13:33:20,539][openai._base_client][INFO] - Retrying request to /chat/completions in 0.803009 seconds
+[2025-12-21 13:33:47,206][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:33:47,413][openai._base_client][INFO] - Retrying request to /chat/completions in 0.469032 seconds
+[2025-12-21 13:33:52,891][openai._base_client][INFO] - Retrying request to /chat/completions in 0.771772 seconds
+[2025-12-21 13:34:06,613][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:34:06,879][openai._base_client][INFO] - Retrying request to /chat/completions in 0.397119 seconds
+[2025-12-21 13:34:14,570][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:34:14,797][openai._base_client][INFO] - Retrying request to /chat/completions in 0.410649 seconds
+[2025-12-21 13:34:20,218][openai._base_client][INFO] - Retrying request to /chat/completions in 0.976470 seconds
+[2025-12-21 13:34:31,312][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:34:31,637][openai._base_client][INFO] - Retrying request to /chat/completions in 0.496590 seconds
+[2025-12-21 13:34:40,039][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:34:40,349][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400080 seconds
+[2025-12-21 13:34:51,600][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:34:51,919][openai._base_client][INFO] - Retrying request to /chat/completions in 0.380585 seconds
+[2025-12-21 13:35:01,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:35:01,720][openai._base_client][INFO] - Retrying request to /chat/completions in 0.428775 seconds
+[2025-12-21 13:35:07,159][openai._base_client][INFO] - Retrying request to /chat/completions in 0.956960 seconds
+[2025-12-21 13:35:16,200][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:35:16,483][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490412 seconds
+[2025-12-21 13:35:24,508][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:35:24,826][openai._base_client][INFO] - Retrying request to /chat/completions in 0.408153 seconds
+[2025-12-21 13:35:30,245][openai._base_client][INFO] - Retrying request to /chat/completions in 0.994561 seconds
+[2025-12-21 13:35:48,088][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:35:48,416][openai._base_client][INFO] - Retrying request to /chat/completions in 0.464616 seconds
+[2025-12-21 13:35:59,847][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:36:00,160][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404818 seconds
+[2025-12-21 13:36:06,510][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:36:06,893][openai._base_client][INFO] - Retrying request to /chat/completions in 0.488868 seconds
+[2025-12-21 13:36:14,967][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:36:15,303][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418218 seconds
+[2025-12-21 13:36:20,737][openai._base_client][INFO] - Retrying request to /chat/completions in 0.772129 seconds
+[2025-12-21 13:36:29,278][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:36:29,600][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481539 seconds
+[2025-12-21 13:36:41,009][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:36:41,242][openai._base_client][INFO] - Retrying request to /chat/completions in 0.478379 seconds
+[2025-12-21 13:36:50,054][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:36:50,418][openai._base_client][INFO] - Retrying request to /chat/completions in 0.497210 seconds
+[2025-12-21 13:36:57,347][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:36:57,679][openai._base_client][INFO] - Retrying request to /chat/completions in 0.481625 seconds
+[2025-12-21 13:37:06,700][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:37:07,026][openai._base_client][INFO] - Retrying request to /chat/completions in 0.447418 seconds
+[2025-12-21 13:37:12,484][openai._base_client][INFO] - Retrying request to /chat/completions in 0.964795 seconds
+[2025-12-21 13:37:22,807][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:37:23,160][openai._base_client][INFO] - Retrying request to /chat/completions in 0.424504 seconds
+[2025-12-21 13:37:30,882][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:37:31,240][openai._base_client][INFO] - Retrying request to /chat/completions in 0.404196 seconds
+[2025-12-21 13:37:36,655][openai._base_client][INFO] - Retrying request to /chat/completions in 0.967872 seconds
+[2025-12-21 13:37:52,674][openai._base_client][INFO] - Retrying request to /chat/completions in 0.435663 seconds
+[2025-12-21 13:37:58,120][openai._base_client][INFO] - Retrying request to /chat/completions in 0.988019 seconds
+[2025-12-21 13:38:08,640][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:38:08,997][openai._base_client][INFO] - Retrying request to /chat/completions in 0.438956 seconds
+[2025-12-21 13:38:30,497][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:38:30,865][openai._base_client][INFO] - Retrying request to /chat/completions in 0.406423 seconds
+[2025-12-21 13:38:40,538][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:38:40,950][openai._base_client][INFO] - Retrying request to /chat/completions in 0.490292 seconds
+[2025-12-21 13:38:49,412][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:38:49,877][openai._base_client][INFO] - Retrying request to /chat/completions in 0.472426 seconds
+[2025-12-21 13:38:55,361][openai._base_client][INFO] - Retrying request to /chat/completions in 0.865031 seconds
+[2025-12-21 13:39:03,910][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:39:04,300][openai._base_client][INFO] - Retrying request to /chat/completions in 0.445240 seconds
+[2025-12-21 13:39:13,230][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:39:13,633][openai._base_client][INFO] - Retrying request to /chat/completions in 0.483987 seconds
+[2025-12-21 13:39:19,127][openai._base_client][INFO] - Retrying request to /chat/completions in 0.798467 seconds
+[2025-12-21 13:39:37,426][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:39:37,810][openai._base_client][INFO] - Retrying request to /chat/completions in 0.409716 seconds
+[2025-12-21 13:39:48,223][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:39:48,576][openai._base_client][INFO] - Retrying request to /chat/completions in 0.391974 seconds
+[2025-12-21 13:39:53,976][openai._base_client][INFO] - Retrying request to /chat/completions in 0.873029 seconds
+[2025-12-21 13:40:02,078][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:40:02,446][openai._base_client][INFO] - Retrying request to /chat/completions in 0.418913 seconds
+[2025-12-21 13:40:11,430][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:40:11,799][openai._base_client][INFO] - Retrying request to /chat/completions in 0.457167 seconds
+[2025-12-21 13:40:23,750][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:40:24,135][openai._base_client][INFO] - Retrying request to /chat/completions in 0.444199 seconds
+[2025-12-21 13:40:32,836][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:40:33,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.443722 seconds
+[2025-12-21 13:40:44,887][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:40:45,301][openai._base_client][INFO] - Retrying request to /chat/completions in 0.440067 seconds
+[2025-12-21 13:40:53,883][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:40:54,207][openai._base_client][INFO] - Retrying request to /chat/completions in 0.475503 seconds
+[2025-12-21 13:41:04,739][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:41:05,075][openai._base_client][INFO] - Retrying request to /chat/completions in 0.400536 seconds
+[2025-12-21 13:41:13,287][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:41:13,707][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63429).
+[2025-12-21 13:41:13,957][openai._base_client][INFO] - Retrying request to /chat/completions in 0.476904 seconds
+[2025-12-21 13:41:23,129][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:41:23,454][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63748).
+[2025-12-21 13:41:23,700][openai._base_client][INFO] - Retrying request to /chat/completions in 0.382367 seconds
+[2025-12-21 13:41:32,165][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:41:32,364][root][WARNING] - Truncated history for env 0 to keep prompt within 64000 tokens (current 63716).
+[2025-12-21 13:41:32,561][openai._base_client][INFO] - Retrying request to /chat/completions in 0.460737 seconds
+[2025-12-21 13:41:38,032][openai._base_client][INFO] - Retrying request to /chat/completions in 0.985943 seconds
diff --git a/outputs/2025-12-21/13-42-37/.hydra/config.yaml b/outputs/2025-12-21/13-42-37/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f8b7ac659f06b114b11649f1a92bfe0a872cccdd
--- /dev/null
+++ b/outputs/2025-12-21/13-42-37/.hydra/config.yaml
@@ -0,0 +1,922 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 400
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 64000
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: null
+  experiment_name: test
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 1
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid. Merge equal tiles
+      by sliding Up, Right, Down, or Left.
+
+      If a move is invalid (no tiles move), a small penalty is applied. Respond with
+      a single action.
+
+      Example: <answer>Up</answer>
+
+      '
+    max_tokens: 256
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 3
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: '0'
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: -1
+  max_turn: 100
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - CoordSokoban
+      n_groups:
+      - 8
+  val:
+    env_groups: 1
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 1
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
+model_config:
+  model_name: gemini-2.5-pro
+  max_concurrency: 16
+model_info:
+  Qwen2.5-7B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-7B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  Qwen2.5-72B-Instruct:
+    provider_name: together
+    model_name: Qwen/Qwen2.5-72B-Instruct-Turbo
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  claude-3.7:
+    provider_name: anthropic
+    model_name: claude-3-7-sonnet-20250219
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-4o-mini:
+    provider_name: openai
+    model_name: gpt-4o-mini
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 512
+  gpt-5-nano-2025-08-07:
+    provider_name: openai
+    model_name: gpt-5-nano-2025-08-07
+    generation_kwargs:
+      max_tokens: 16342
+  deepseek-r1:
+    provider_name: deepseek
+    model_name: deepseek-reasoner
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  ark-deepseek-v3-250324:
+    provider_name: openai
+    model_name: ark-deepseek-v3-250324
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  deepseek-v3:
+    provider_name: deepseek
+    model_name: deepseek-chat
+    generation_kwargs:
+      temperature: 0
+      max_completion_tokens: 512
+  gemini-2.5-pro:
+    provider_name: openai
+    model_name: gemini-2.5-pro
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 256
+  TA/openai/gpt-oss-120b:
+    provider_name: openai
+    model_name: TA/openai/gpt-oss-120b
+    generation_kwargs:
+      temperature: 0
+      max_tokens: 16342
+rollout:
+  max_model_len: 64000
diff --git a/outputs/2025-12-21/13-42-37/eval_api.log b/outputs/2025-12-21/13-42-37/eval_api.log
new file mode 100644
index 0000000000000000000000000000000000000000..bbcce033446d4fc8288a4116738be62d14776e84
--- /dev/null
+++ b/outputs/2025-12-21/13-42-37/eval_api.log
@@ -0,0 +1,13 @@
+[2025-12-21 13:42:44,779][openai._base_client][INFO] - Retrying request to /chat/completions in 0.489042 seconds
+[2025-12-21 13:42:48,988][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:42:54,002][openai._base_client][INFO] - Retrying request to /chat/completions in 0.387857 seconds
+[2025-12-21 13:42:58,861][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:43:03,871][openai._base_client][INFO] - Retrying request to /chat/completions in 0.466027 seconds
+[2025-12-21 13:43:08,879][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:43:13,891][openai._base_client][INFO] - Retrying request to /chat/completions in 0.455945 seconds
+[2025-12-21 13:43:18,191][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:43:23,200][openai._base_client][INFO] - Retrying request to /chat/completions in 0.420455 seconds
+[2025-12-21 13:43:28,293][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:43:33,304][openai._base_client][INFO] - Retrying request to /chat/completions in 0.422127 seconds
+[2025-12-21 13:43:37,242][httpx][INFO] - HTTP Request: POST https://api.ohmygpt.com/v1/chat/completions "HTTP/1.1 200 OK"
+[2025-12-21 13:43:42,256][openai._base_client][INFO] - Retrying request to /chat/completions in 0.468679 seconds
diff --git a/outputs/2025-12-21/19-57-54/.hydra/hydra.yaml b/outputs/2025-12-21/19-57-54/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..fd432e7d4885d0f7ed663f59867694f1a3a4969b
--- /dev/null
+++ b/outputs/2025-12-21/19-57-54/.hydra/hydra.yaml
@@ -0,0 +1,209 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+    - 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4
+      grid.\nMerge equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed
+      grid where (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid
+      (no tiles move), a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+    - actor_rollout_ref.rollout.max_model_len=14400
+    - agent_proxy.max_turn=700
+    - actor_rollout_ref.rollout.gpu_memory_utilization=0.5
+    - trainer.experiment_name=_2048_starpo_stateestimation
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: 'actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.gpu_memory_utilization=0.5,actor_rollout_ref.rollout.max_model_len=14400,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,agent_proxy.max_turn=700,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,custom_envs.game_2048.env_instruction="You
+      are playing the 2048 game on a 4x4 grid.\nMerge equal tiles by sliding Up, Right,
+      Down, or Left.\nUse a zero-indexed grid where (0,0) is top-left and (3,3) is
+      bottom-right.\nIf a move is invalid (no tiles move), a small penalty is applied.\n\nExample
+      answer format:\n<think>\n<observation>\n[Current Grid Symbol Representation]\nTile
+      values and positions: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nI
+      want to merge the two 2s in the top row. Sliding Right will merge them at (0,3).\n<prediction>\n[Predicted
+      Grid Symbol Representation after move]\nPredicted changes: The 2 at (0,0) and
+      2 at (0,1) merge into a 4 at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>",model_path=/mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct,system.CUDA_VISIBLE_DEVICES=''0,1,2,3,4,5,6,7'',trainer.experiment_name=_2048_starpo_stateestimation,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it'
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/19-57-54
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/20-03-21/.hydra/config.yaml b/outputs/2025-12-21/20-03-21/.hydra/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..900aafb6b37d3371157b7fa6fae5d45f45262d79
--- /dev/null
+++ b/outputs/2025-12-21/20-03-21/.hydra/config.yaml
@@ -0,0 +1,862 @@
+actor_rollout_ref:
+  actor:
+    optim:
+      _target_: verl.workers.config.FSDPOptimizerConfig
+      lr: 1.0e-06
+      lr_warmup_steps_ratio: 0.0
+      total_training_steps: -1
+      weight_decay: 0.01
+      lr_warmup_steps: -1
+      betas:
+      - 0.9
+      - 0.999
+      clip_grad: 1.0
+      min_lr_ratio: 0.0
+      num_cycles: 0.5
+      warmup_style: constant
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    _target_: verl.workers.config.FSDPActorConfig
+    strategy: fsdp
+    ppo_mini_batch_size: ${ppo_mini_batch_size}
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.28
+    freeze_vision_tower: false
+    policy_loss:
+      _target_: verl.workers.config.PolicyLossConfig
+      loss_mode: vanilla
+      clip_cov_ratio: 0.0002
+      clip_cov_lb: 1.0
+      clip_cov_ub: 5.0
+      kl_cov_ratio: 0.0002
+      ppo_kl_coef: 0.1
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    tis_imp_ratio_cap: -1
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.0
+    kl_loss_type: kl
+    ppo_epochs: 1
+    shuffle: false
+    checkpoint:
+      _target_: verl.trainer.config.CheckpointConfig
+      save_contents:
+      - model
+      - optimizer
+      - extra
+      load_contents: ${.save_contents}
+      async_save: false
+    use_fused_kernels: ${oc.select:actor_rollout_ref.model.use_fused_kernels,false}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    grad_clip: 1.0
+    ulysses_sequence_parallel_size: 1
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+    use_remove_padding: ${oc.select:actor_rollout_ref.model.use_remove_padding,false}
+    use_ref: true
+    grpo_advantage_length_weight: ${grpo_advantage_length_weight}
+  ref:
+    strategy: ${actor_rollout_ref.actor.strategy}
+    use_torch_compile: ${oc.select:actor_rollout_ref.actor.use_torch_compile,true}
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: false
+      all_ranks: false
+      ranks: []
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config:
+        nsys:
+          _target_: verl.utils.profiler.config.NsightToolConfig
+          discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+        npu:
+          _target_: verl.utils.profiler.config.NPUToolConfig
+          contents: []
+          level: level1
+          analysis: true
+          discrete: false
+        torch:
+          _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+          step_start: 0
+          step_end: null
+        torch_memory:
+          _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+          trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+          stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    model: null
+    ulysses_sequence_parallel_size: ${oc.select:actor_rollout_ref.actor.ulysses_sequence_parallel_size,1}
+    entropy_from_logits_with_chunking: false
+    entropy_checkpointing: false
+  rollout:
+    _target_: verl.workers.config.RolloutConfig
+    name: vllm
+    mode: sync
+    temperature: 1
+    top_k: -1
+    top_p: 1
+    prompt_length: 1
+    response_length: 256
+    dtype: bfloat16
+    gpu_memory_utilization: 0.5
+    ignore_eos: false
+    enforce_eager: true
+    cudagraph_capture_sizes: null
+    free_cache_engine: true
+    tensor_model_parallel_size: 1
+    data_parallel_size: 1
+    expert_parallel_size: 1
+    max_num_batched_tokens: 8192
+    max_model_len: 14400
+    max_num_seqs: 1024
+    enable_chunked_prefill: true
+    enable_prefix_caching: true
+    load_format: dummy
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+    log_prob_use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+    log_prob_max_token_len_per_gpu: ${oc.select:actor_rollout_ref.actor.ppo_max_token_len_per_gpu,16384}
+    disable_log_stats: true
+    do_sample: true
+    'n': 1
+    over_sample_rate: 0
+    multi_stage_wake_up: false
+    engine_kwargs:
+      vllm: {}
+      sglang: {}
+    val_kwargs:
+      _target_: verl.workers.config.SamplingConfig
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.5
+      'n': 1
+      do_sample: true
+    multi_turn:
+      _target_: verl.workers.config.MultiTurnConfig
+      enable: false
+      max_assistant_turns: null
+      tool_config_path: null
+      max_user_turns: null
+      max_parallel_calls: 1
+      max_tool_response_length: 256
+      tool_response_truncate_side: middle
+      interaction_config_path: null
+      use_inference_chat_template: false
+      tokenization_sanity_check_mode: strict
+      format: hermes
+      num_repeat_rollouts: null
+    calculate_log_probs: false
+    agent:
+      _target_: verl.workers.config.AgentLoopConfig
+      num_workers: 8
+      agent_loop_config_path: null
+      custom_async_server:
+        _target_: verl.workers.config.CustomAsyncServerConfig
+        path: null
+        name: null
+    update_weights_bucket_megabytes: 512
+    trace:
+      _target_: verl.workers.config.TraceConfig
+      backend: null
+      token2text: false
+    skip_rollout: false
+    skip_dump_dir: /tmp/rollout_dump
+    skip_tokenizer_init: true
+    profiler:
+      _target_: verl.utils.profiler.ProfilerConfig
+      tool: ${oc.select:global_profiler.tool,null}
+      enable: ${oc.select:actor_rollout_ref.actor.profiler.enable,false}
+      all_ranks: ${oc.select:actor_rollout_ref.actor.profiler.all_ranks,false}
+      ranks: ${oc.select:actor_rollout_ref.actor.profiler.ranks,[]}
+      save_path: ${oc.select:global_profiler.save_path,null}
+      tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+    layered_summon: false
+    rollout_filter_ratio: 0.25
+    rollout_filter_type: largest
+    rollout_filter_metric: reward_variance
+  model:
+    _target_: verl.workers.config.HFModelConfig
+    path: ${model_path}
+    hf_config_path: null
+    tokenizer_path: null
+    use_shm: false
+    trust_remote_code: false
+    custom_chat_template: null
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+    exclude_modules: null
+    use_liger: false
+    use_fused_kernels: false
+    fused_kernel_options:
+      impl_backend: torch
+  hybrid_engine: true
+  nccl_timeout: 600
+data:
+  tokenizer: null
+  use_shm: false
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: null
+  max_response_length: null
+  train_batch_size: null
+  val_batch_size: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  return_full_prompt: false
+  shuffle: true
+  dataloader_num_workers: 8
+  validation_shuffle: false
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  image_key: images
+  video_key: videos
+  trust_remote_code: false
+  custom_cls:
+    path: null
+    name: null
+  return_multi_modal_inputs: true
+  sampler:
+    class_path: null
+    class_name: null
+  datagen:
+    path: null
+    name: null
+  apply_chat_template_kwargs: {}
+critic:
+  optim:
+    _target_: verl.workers.config.FSDPOptimizerConfig
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    total_training_steps: -1
+    weight_decay: 0.01
+    lr_warmup_steps: -1
+    betas:
+    - 0.9
+    - 0.999
+    clip_grad: 1.0
+    min_lr_ratio: 0.0
+    num_cycles: 0.5
+    warmup_style: constant
+  model:
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      offload_policy: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+      model_dtype: fp32
+      use_orig_params: false
+      ulysses_sequence_parallel_size: 1
+      entropy_from_logits_with_chunking: false
+      use_torch_compile: true
+      entropy_checkpointing: false
+      forward_only: false
+      strategy: fsdp
+    path: ${model_path}
+    tokenizer_path: ${oc.select:actor_rollout_ref.model.path,"~/models/deepseek-llm-7b-chat"}
+    override_config: {}
+    external_lib: ${oc.select:actor_rollout_ref.model.external_lib,null}
+    trust_remote_code: ${oc.select:actor_rollout_ref.model.trust_remote_code,false}
+    _target_: verl.workers.config.FSDPCriticModelCfg
+    use_shm: false
+    enable_gradient_checkpointing: true
+    enable_activation_offload: false
+    use_remove_padding: false
+    lora_rank: ${lora.rank}
+    lora_alpha: ${lora.alpha}
+    target_modules: ${lora.target_modules}
+  _target_: verl.workers.config.FSDPCriticConfig
+  rollout_n: ${oc.select:actor_rollout_ref.rollout.n,1}
+  strategy: fsdp
+  enable: null
+  ppo_mini_batch_size: ${ppo_mini_batch_size}
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: ${micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${oc.select:actor_rollout_ref.actor.use_dynamic_bsz,false}
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: ${.ppo_max_token_len_per_gpu}
+  ppo_epochs: ${oc.select:actor_rollout_ref.actor.ppo_epochs,1}
+  shuffle: ${oc.select:actor_rollout_ref.actor.shuffle,false}
+  cliprange_value: 0.5
+  loss_agg_mode: ${oc.select:actor_rollout_ref.actor.loss_agg_mode,token-mean}
+  checkpoint:
+    _target_: verl.trainer.config.CheckpointConfig
+    save_contents:
+    - model
+    - optimizer
+    - extra
+    load_contents: ${.save_contents}
+    async_save: false
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config:
+      nsys:
+        _target_: verl.utils.profiler.config.NsightToolConfig
+        discrete: ${oc.select:global_profiler.global_tool_config.nsys.discrete}
+      npu:
+        _target_: verl.utils.profiler.config.NPUToolConfig
+        contents: []
+        level: level1
+        analysis: true
+        discrete: false
+      torch:
+        _target_: verl.utils.profiler.config.TorchProfilerToolConfig
+        step_start: 0
+        step_end: null
+      torch_memory:
+        _target_: verl.utils.profiler.config.TorchMemoryToolConfig
+        trace_alloc_max_entries: ${oc.select:global_profiler.global_tool_config.torch_memory.trace_alloc_max_entries,100000}
+        stack_depth: ${oc.select:global_profiler.global_tool_config.torch_memory.stack_depth,32}
+  forward_micro_batch_size: ${oc.select:.ppo_micro_batch_size,null}
+  forward_micro_batch_size_per_gpu: ${oc.select:.ppo_micro_batch_size_per_gpu,null}
+  ulysses_sequence_parallel_size: 1
+  grad_clip: 1.0
+reward_model:
+  enable: false
+  enable_resource_pool: false
+  n_gpus_per_node: 0
+  nnodes: 0
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    trust_remote_code: false
+    use_shm: false
+    use_remove_padding: false
+    use_fused_kernels: ${actor_rollout_ref.model.use_fused_kernels}
+    fsdp_config:
+      _target_: verl.workers.config.FSDPEngineConfig
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      reshard_after_forward: true
+      fsdp_size: -1
+      forward_prefetch: false
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+  reward_manager: naive
+  launch_reward_fn_async: false
+  sandbox_fusion:
+    url: null
+    max_concurrent: 64
+    memory_limit_mb: 1024
+  profiler:
+    _target_: verl.utils.profiler.ProfilerConfig
+    tool: ${oc.select:global_profiler.tool,null}
+    enable: false
+    all_ranks: false
+    ranks: []
+    save_path: ${oc.select:global_profiler.save_path,null}
+    tool_config: ${oc.select:actor_rollout_ref.actor.profiler.tool_config,null}
+  ulysses_sequence_parallel_size: 1
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  _target_: verl.trainer.config.AlgoConfig
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  norm_adv_by_std_in_grpo: true
+  use_kl_in_reward: false
+  kl_penalty: kl
+  kl_ctrl:
+    _target_: verl.trainer.config.KLControlConfig
+    type: fixed
+    kl_coef: 0.0
+    horizon: 10000
+    target_kl: 0.1
+  use_pf_ppo: false
+  pf_ppo:
+    reweight_method: pow
+    weight_pow: 2.0
+  high_level_gamma: 0.95
+  bi_level_gae: false
+trainer:
+  balance_batch: true
+  total_epochs: 30
+  total_training_steps: 200
+  project_name: ragen_latest_qwen_05B_it
+  experiment_name: _2048_starpo_stateestimation
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  rollout_data_dir: null
+  validation_data_dir: null
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  esi_redundant_time: 0
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: true
+  val_only: false
+  test_freq: 10
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: /mnt/general/wanghy/RAGEN/saves/
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  device: cuda
+  use_legacy_worker_impl: auto
+  local_log_dir: results/
+  validation_steps: 1
+  generations_to_log_to_wandb:
+    train: 128
+    val: 20
+global_profiler:
+  _target_: verl.utils.profiler.ProfilerConfig
+  tool: null
+  steps: null
+  profile_continuous_steps: false
+  save_path: outputs/profile
+  global_tool_config:
+    nsys:
+      _target_: verl.utils.profiler.config.NsightToolConfig
+      discrete: false
+      controller_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+      worker_nsight_options:
+        trace: cuda,nvtx,cublas,ucx
+        cuda-memory-usage: 'true'
+        cuda-graph-trace: graph
+        capture-range: cudaProfilerApi
+        capture-range-end: null
+        kill: none
+    torch_memory:
+      trace_alloc_max_entries: 100000
+      stack_depth: 32
+      context: all
+      stacks: all
+      kw_args: {}
+ray_kwargs:
+  ray_init:
+    num_cpus: null
+  timeline_json_file: null
+custom_envs:
+  SimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+  LargerSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      dim_x: 8
+      dim_y: 8
+      num_boxes: 2
+      max_steps: 100
+      search_depth: 10
+  SokobanDifferentGridVocab:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. \nYou are the player and\
+      \ you need to push all boxes to targets. \nWhen you are right next to a box,\
+      \ you can push it by moving in the same direction. \nYou cannot push a box through\
+      \ a wall, and you cannot pull a box. \nThe answer should be a sequence of actions,\
+      \ like <answer>Right || Right || Up</answer>\n"
+    max_tokens: 100
+    env_config:
+      search_depth: 30
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      grid_lookup:
+        0: W
+        1: .
+        2: G
+        3: C
+        4: B
+        5: A
+        6: '@'
+      grid_vocab:
+        W: wall
+        .: empty
+        G: target
+        C: box on target
+        B: box
+        A: player
+        '@': player on target
+  CoordSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: "You are solving the Sokoban puzzle. You are the player and you\
+      \ need to push all boxes to targets.\nYou are provided with a symbol grid and\
+      \ the zero-indexed coordinates of the player, each box, and each target. \n\
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner\
+      \ (5, 5). \nWhen you are exactly next to a box, you can push it by moving in\
+      \ the same direction. \nYou cannot push a box through a wall, and you cannot\
+      \ pull a box.\nThe answer should be a sequence of actions, like <answer>Right\
+      \ || Right || Up</answer>.\n"
+    max_tokens: 120
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 2
+      max_steps: 100
+      observation_format: grid_coord
+  VisualSimpleSokoban:
+    env_type: sokoban
+    max_actions_per_traj: 10
+    env_instruction: You are solving the Sokoban puzzle. You are the player and you
+      need to push all boxes to targets. When you are right next to a box, you can
+      push it by moving in the same direction. You cannot push a box through a wall,
+      and you cannot pull a box. The answer should be a sequence of actions, like
+      <answer>Right || Right || Up</answer>
+    max_tokens: 100
+    env_config:
+      dim_x: 6
+      dim_y: 6
+      num_boxes: 1
+      max_steps: 100
+      render_mode: rgb_array
+  Countdown:
+    env_type: countdown
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving the Countdown puzzle. You should use the num
+      list to create an equation that equals the target. Example answer format: <think>
+      To find an equation using [3, 5, 2] to get 4. Let''s check 2 + 5 = 7, 7 - 3
+      = 4. So the answer is 2 + 5 - 3 = 4. </think><answer>2 + 5 - 3</answer>'
+    max_tokens: 100
+    env_config: null
+  Bandit:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: train
+  BanditTest:
+    env_type: bandit
+    max_actions_per_traj: 1
+    env_instruction: ''
+    max_tokens: 100
+    env_config:
+      split: test
+  FrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. Forbid the whole and
+      go to the target. You may move to the unintended direction due to the slippery
+      ice. Example answer format: <think>To forbid the hole and go to the target,
+      I should go left then go up.</think><answer>Left || Up</answer>'
+    max_tokens: 100
+    env_config: null
+  CoordFrozenLake:
+    env_type: frozen_lake
+    max_actions_per_traj: 25
+    env_instruction: 'You are solving the FrozenLake puzzle. The observation includes
+      both a symbol grid and zero-indexed coordinates for the start, goal, player,
+      and any holes.
+
+      Coordinates range from the top-left corner (0, 0) to the bottom-right corner
+      (5, 5).
+
+      Beware that the ice is slippery, so the agent might slide and end up in an unintended
+      tile.
+
+      Respond with a sequence of actions such as <answer>Left || Up || Up</answer>.
+
+      '
+    max_tokens: 120
+    env_config:
+      observation_format: grid_coord
+  MetamathQA:
+    env_type: metamathqa
+    max_actions_per_traj: 1
+    env_instruction: 'You are solving Math problems. '
+    max_tokens: 100
+    env_config: null
+  WebShop:
+    env_type: webshop
+    max_actions_per_traj: 9
+    env_instruction: 'You are browsing an online shop. Based on the instruction, buy
+      a product that close to the production description. You need to search, read
+      the search results, pick a product, choose the size and color and buy. You should
+      only choose action from the available actions list provided later.  Example
+      process: I need a gingko light and 20x20 pillow cover that is hand painted.
+      First search[gingko light 20x20 pillow cover hand painted], answer format: <answer>search[blanket
+      with fleece throw]</answer>. Valid answer is search[<keywords>] or click[<clickable>].'
+    max_tokens: 200
+    env_config:
+      dataset: small
+  Lean:
+    env_type: lean
+    max_actions_per_traj: 30
+    env_instruction: You are a Lean theorem prover. Given a Lean theorem statement,
+      propose a sequence of tactics that completes the proof. Think step by step about
+      which tactics to apply next. Provide tactics separated by '||', for example
+      <answer>intro || simp || rfl</answer>.
+    max_tokens: 512
+    env_config: null
+  game_2048:
+    env_type: game_2048
+    max_actions_per_traj: 700
+    env_instruction: 'You are playing the 2048 game on a 4x4 grid.\nMerge equal tiles
+      by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where (0,0) is
+      top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move), a
+      small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+      Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+      4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+      Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+      after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4
+      at (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA
+      sample full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n.
+      . . .\n. . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+      moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge
+      to the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile
+      4 created at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>'
+    max_tokens: 8192
+    env_config: null
+  blackjack:
+    env_type: blackjack
+    max_actions_per_traj: 10
+    env_instruction: 'You are playing Blackjack against a dealer. The dealer must
+      hit on 16 or less and stand on 17 or more.
+
+      Choose either Stick or Hit. Respond with a single action.
+
+      Example: <answer>Hit</answer>
+
+      '
+    max_tokens: 64
+    env_config: null
+  rubikscube:
+    env_type: rubikscube
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a 2x2 Rubik''s Cube (Pocket Cube). The goal
+      is to restore the cube so that each of the faces consists of a single, unique
+      color.
+
+      Available actions use standard Singmaster notation for face rotations: U, U'',
+      D, D'', L, L'', R, R'', F, F'', B, B''.
+
+      - Faces: U (Up), D (Down), L (Left), R (Right), F (Front), B (Back).
+
+      - Modifiers: A letter alone means 90° clockwise (e.g., ''R''). A letter with
+      prime ('') means 90° counter-clockwise (e.g., "R''").
+
+      Respond with a sequence of actions separated by "||".
+
+      Example: <answer>U</answer>
+
+      '
+    max_tokens: 96
+    env_config:
+      scramble_depth: 1
+      max_steps: 20
+      render_mode: text
+  SimpleSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 20
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 2x2 box contains the numbers 1-4 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format, for example: <answer>place 1 at
+      row 2 col 3</answer> or <answer>1,2,3</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 4
+      difficulty: easy
+      render_format: with_feedback
+      show_conflicts: false
+      show_valid_numbers: false
+      max_steps: 20
+  MediumSudoku:
+    env_type: sudoku
+    max_actions_per_traj: 30
+    env_instruction: 'You are solving a Sudoku puzzle. Fill in the grid so that every
+      row, column, and 3x3 box contains the numbers 1-9 without repetition.
+
+      Initial cells are shown in [brackets] and cannot be modified. Empty cells are
+      shown as dots (.).
+
+      Place numbers one at a time using the format: <answer>place 5 at row 2 col 3</answer>
+      or <answer>2,3,5</answer>
+
+      The environment will provide feedback on valid/invalid moves and show conflicts
+      if any occur.
+
+      '
+    max_tokens: 150
+    parallel_friendly: false
+    max_workers: 32
+    env_config:
+      grid_size: 9
+      difficulty: medium
+      render_format: with_feedback
+      show_conflicts: true
+      show_valid_numbers: true
+      max_steps: 81
+system:
+  CUDA_VISIBLE_DEVICES: 0,1,2,3,4,5,6,7
+seed:
+  train: 10000
+  val: 123
+micro_batch_size_per_gpu: 1
+ppo_mini_batch_size: 32
+model_path: /mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+enable_response_mask: true
+grpo_advantage_length_weight: false
+lora:
+  rank: 0
+  alpha: 16
+  target_modules: all-linear
+agent_proxy:
+  max_context_window: 5
+  max_turn: 700
+  action_sep: '||'
+  max_actions_per_turn: 1
+  use_turn_scores: false
+  enable_think: true
+  reward_normalization:
+    grouping: state
+    method: identity
+es_manager:
+  format_penalty: -0.1
+  train:
+    env_groups: 8
+    group_size: 16
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 8
+  val:
+    env_groups: 256
+    group_size: 1
+    env_configs:
+      tags:
+      - game_2048
+      n_groups:
+      - 256
+ctx_manager:
+  generation:
+    gen_config:
+      response_length: ${actor_rollout_ref.rollout.response_length}
+      temperature: ${actor_rollout_ref.rollout.temperature}
+      top_p: ${actor_rollout_ref.rollout.top_p}
+      top_k: ${actor_rollout_ref.rollout.top_k}
+      kwargs: null
diff --git a/outputs/2025-12-21/20-03-21/.hydra/overrides.yaml b/outputs/2025-12-21/20-03-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f2c159515cd0e60b2050f9209fe97293334c9377
--- /dev/null
+++ b/outputs/2025-12-21/20-03-21/.hydra/overrides.yaml
@@ -0,0 +1,27 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+- 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4 grid.\nMerge
+  equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where
+  (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move),
+  a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+  Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+  4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+  Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+  after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4 at
+  (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA sample
+  full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n. . . .\n.
+  . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+  moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge to
+  the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile 4 created
+  at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+- actor_rollout_ref.rollout.max_model_len=14400
+- agent_proxy.max_turn=700
+- actor_rollout_ref.rollout.gpu_memory_utilization=0.5
+- trainer.experiment_name=_2048_starpo_stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-21/20-05-21/.hydra/hydra.yaml b/outputs/2025-12-21/20-05-21/.hydra/hydra.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..66e6cc91e8e6b5f0ab04059a196fcd49623fe058
--- /dev/null
+++ b/outputs/2025-12-21/20-05-21/.hydra/hydra.yaml
@@ -0,0 +1,179 @@
+hydra:
+  run:
+    dir: outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+    - trainer.project_name=ragen_latest_qwen_05B_it
+    - trainer.n_gpus_per_node=8
+    - model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+    - trainer.experiment_name=game_2048-frommlp-windows5
+    - algorithm.adv_estimator=gae
+    - algorithm.kl_ctrl.kl_coef=0.0
+    - actor_rollout_ref.actor.kl_loss_coef=0.0
+    - actor_rollout_ref.actor.clip_ratio_high=0.28
+    - actor_rollout_ref.rollout.rollout_filter_ratio=0.25
+  job:
+    name: train
+    chdir: null
+    override_dirname: actor_rollout_ref.actor.clip_ratio_high=0.28,actor_rollout_ref.actor.kl_loss_coef=0.0,actor_rollout_ref.rollout.rollout_filter_ratio=0.25,algorithm.adv_estimator=gae,algorithm.kl_ctrl.kl_coef=0.0,model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5,system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7',trainer.experiment_name=game_2048-frommlp-windows5,trainer.n_gpus_per_node=8,trainer.project_name=ragen_latest_qwen_05B_it
+    id: ???
+    num: ???
+    config_name: _8_2048
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /mnt/general/wanghy/RAGEN
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /mnt/general/wanghy/RAGEN/config
+      schema: file
+      provider: main
+    - path: /mnt/general/wanghy/RAGEN/verl/verl/trainer/config
+      schema: file
+      provider: command-line
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /mnt/general/wanghy/RAGEN/outputs/2025-12-21/20-05-21
+    choices:
+      reward_model: dp_reward_model
+      critic: dp_critic
+      critic/../engine@critic.model.fsdp_config: fsdp
+      critic/../optim@critic.optim: fsdp
+      model@actor_rollout_ref.model: hf_model
+      rollout@actor_rollout_ref.rollout: rollout
+      ref@actor_rollout_ref.ref: dp_ref
+      ref/../engine@actor_rollout_ref.ref.fsdp_config: fsdp
+      data: legacy_data
+      actor@actor_rollout_ref.actor: dp_actor
+      actor/../engine@actor_rollout_ref.actor.fsdp_config: fsdp
+      actor/../optim@actor_rollout_ref.actor.optim: fsdp
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false
diff --git a/outputs/2025-12-21/20-05-21/.hydra/overrides.yaml b/outputs/2025-12-21/20-05-21/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..0fc63e1a59e022e800bc29ca7368187127fcc701
--- /dev/null
+++ b/outputs/2025-12-21/20-05-21/.hydra/overrides.yaml
@@ -0,0 +1,10 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/wanghy/LLaMA-Factory/saves/qwen3/full/sft/qwen2.5_0.5B_it_2048_windows5
+- trainer.experiment_name=game_2048-frommlp-windows5
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25
diff --git a/outputs/2025-12-21/20-05-21/train.log b/outputs/2025-12-21/20-05-21/train.log
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/outputs/2025-12-21/20-09-41/.hydra/overrides.yaml b/outputs/2025-12-21/20-09-41/.hydra/overrides.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..8c20842a1da13eaff0d502c5bdc417dede35b97d
--- /dev/null
+++ b/outputs/2025-12-21/20-09-41/.hydra/overrides.yaml
@@ -0,0 +1,26 @@
+- system.CUDA_VISIBLE_DEVICES='0,1,2,3,4,5,6,7'
+- trainer.project_name=ragen_latest_qwen_05B_it
+- trainer.n_gpus_per_node=8
+- model_path=/mnt/general/share/model/Qwen/Qwen2.5-0.5B-Instruct
+- 'custom_envs.game_2048.env_instruction="You are playing the 2048 game on a 4x4 grid.\nMerge
+  equal tiles by sliding Up, Right, Down, or Left.\nUse a zero-indexed grid where
+  (0,0) is top-left and (3,3) is bottom-right.\nIf a move is invalid (no tiles move),
+  a small penalty is applied.\n\nExample answer format:\n<think>\n<observation>\n[Current
+  Grid Symbol Representation]\nTile values and positions: 2 at (0,0), 2 at (0,1),
+  4 at (3,3).\n</observation>\nI want to merge the two 2s in the top row. Sliding
+  Right will merge them at (0,3).\n<prediction>\n[Predicted Grid Symbol Representation
+  after move]\nPredicted changes: The 2 at (0,0) and 2 at (0,1) merge into a 4 at
+  (0,3). (3,3) remains 4.\n</prediction>\n</think>\n<answer>Right</answer>\n\nA sample
+  full output is as follows:\n<think>\n<observation>\n2 2 . .\n. . . .\n. . . .\n.
+  . . 4\nNon-empty tiles: 2 at (0,0), 2 at (0,1), 4 at (3,3).\n</observation>\nChecking
+  moves:\nUp: No merge.\nRight: (0,0) and (0,1) are equal (2), they will merge to
+  the right.\n<prediction>\n. . . 4\n. . . .\n. . . .\n. . . 4\nMerged tile 4 created
+  at (0,3). Tile at (3,3) stays.\n</prediction>\n</think>\n<answer>Right</answer>"'
+- actor_rollout_ref.rollout.max_model_len=14400
+- agent_proxy.max_turn=700
+- trainer.experiment_name=_2048_starpo_stateestimation
+- algorithm.adv_estimator=gae
+- algorithm.kl_ctrl.kl_coef=0.0
+- actor_rollout_ref.actor.kl_loss_coef=0.0
+- actor_rollout_ref.actor.clip_ratio_high=0.28
+- actor_rollout_ref.rollout.rollout_filter_ratio=0.25